Artificial Intelligence Review (2024) 57:296

https://doi.org/10.1007/s10462-024-10906-z

A systematic review of aspect‑based sentiment analysis:

domains, methods, and trends

Yan Cathy Hua1 · Paul Denny1 · Jörg Wicker1 · Katerina Taskova1

Accepted: 6 August 2024 / Published online: 17 September 2024

© The Author(s) 2024

Abstract
Aspect-based sentiment analysis (ABSA) is a fine-grained type of sentiment analysis
that identifies aspects and their associated opinions from a given text. With the surge of
digital opinionated text data, ABSA gained increasing popularity for its ability to mine
more detailed and targeted insights. Many review papers on ABSA subtasks and solution
methodologies exist, however, few focus on trends over time or systemic issues relating to
research application domains, datasets, and solution approaches. To fill the gap, this paper
presents a systematic literature review (SLR) of ABSA studies with a focus on trends and
high-level relationships among these fundamental components. This review is one of the
largest SLRs on ABSA. To our knowledge, it is also the first to systematically examine the
interrelations among ABSA research and data distribution across domains, as well as trends
in solution paradigms and approaches. Our sample includes 727 primary studies screened
from 8550 search results without time constraints via an innovative automatic filtering pro-
cess. Our quantitative analysis not only identifies trends in nearly two decades of ABSA
research development but also unveils a systemic lack of dataset and domain diversity as
well as domain mismatch that may hinder the development of future ABSA research. We
discuss these findings and their implications and propose suggestions for future research.

Keywords ABSA · Aspect-based sentiment analysis · Systematic literature review · Natural

language processing

* Yan Cathy Hua

[email protected]
Paul Denny
[email protected]
Jörg Wicker
[email protected]
Katerina Taskova
[email protected]
1
School of Computer Science, The University of Auckland, Auckland, New Zealand

Vol.:(0123456789)
296 Page 2 of 51 Y. C. Hua et al.

1 Introduction

In the digital era, a vast amount of online opinionated text is generated daily through
which people express views and feelings (i.e. sentiment) towards certain subjects, such
as user reviews, social media posts, and open-ended survey question responses (Kumar
and Gupta 2021). Understanding the sentiment of these opinionated text data is essential
for gaining insights into people’s preferences and behaviours and supporting decision-
making across a wide variety of domains (Sharma and Shekhar 2020; Wankhade et al
2022; Tubishat et al 2021; García-Pablos et al 2018; Poria et al 2016). The analyses of
opinionated text usually aim at answering questions such as “What subjects were men-
tioned?”, “What did people think of (a specific subject)?”, and “How are the subjects
and/or opinions distributed across the sample?” (e.g. (Dragoni et al 2019; Krishnaku-
mari and Sivasankar 2018; Fukumoto et al 2016; Zarindast et al 2021)). These objec-
tives, along with today’s enormous volume of digital opinionated text, require an
automated solution for identifying, extracting and classifying the subjects and their
associated opinions from the raw text. Aspect-based sentiment analysis (ABSA) is one
such solution.

1.1 Review focus and research questions

This work presents a systematic literature review (SLR) of existing ABSA studies with
a large-scale sample and quantitative results. We focus on trends and high-level patterns
instead of methodological details that were well covered by the existing surveys mentioned
above. We aim to benefit both ABSA newcomers by introducing the basics of the topic, as
well as existing ABSA researchers by sharing perspectives and findings that are useful to
the ABSA community and can only be obtained beyond the immediate research tasks and
technicalities.
We seek to answer the following sets of research questions (RQs):

RQ1. To what extent is ABSA research and its dataset resources dominated by the com-
mercial (especially the product and service review) domain? What proportion of ABSA
research focuses on other domains and dataset resources?
RQ2. What are the most common ABSA problem formulations via subtask combina-
tions, and what proportion of ABSA studies only focus on a specific subtask?
RQ3. What is the trend in the ABSA solution approaches over time? Are linguistic and
traditional machine-learning approaches still in use?

This review makes a number of unique contributions to the ABSA research field: (1) It
is one of the largest scoped SLRs on ABSA, with a main review and a Phase-2 targeted
review of a combined 727 primary studies published in 2008–2024, selected from 8550
search results without time constraint. (2) To our knowledge, it is the first SLR that sys-
tematically examines the ABSA data resource distribution in relation to research applica-
tion domains and methodologies; and (3) Our review methodology adopted an innovative
automatic filtering process based on PDF-mining, which enhanced screening quality and
reliability. Our quantitative results not only revealed trends in nearly two decades of ABSA
research literature but also highlighted potential systemic issues that could limit the devel-
opment of future ABSA research.
A systematic review of aspect‑based sentiment analysis: domains,… Page 3 of 51 296

1.2 Organisation of this review

In Sect. 2 (“Background”), we introduce ABSA and highlight the motivation and unique-
ness of this review. Section 3 (“Methods”) outlines our SLR procedures, and Sect. 4
(“Results”) answers the research questions with the SLR results. We then discuss the key
findings and acknowledge limitations in Sects. 5 and 6 (“Discussion” and “Conclusion”).
For those interested in more details, Appendix A provides an in-depth introduction to
ABSA and its subtasks. Appendix B describes the full details of our Methods, and addi-
tional figures from the Results are provided in Appendix C.

2 Background

2.1 ABSA: a fine‑grained sentiment analysis

Aspect-based sentiment analysis (ABSA) is a sub-field of Sentiment Analysis (SA), which

is a core task of natural language processing (NLP). SA, also known as “opinion mining”
(García-Pablos et al 2018; Poria et al 2016; Liang et al 2022; López and Arco 2019; Tran
et al 2020), solves the problem of identifying and classifying given text corpora’s affect or
sentiment orientation (Akhtar et al 2020; Tubishat et al 2021) into polarity categories (e.g.
“positive, neutral, negative”) (Brauwers and Frasincar 2023; Hu and Liu 2004a), intensity/
strength scores (e.g. from 1 to 5) (Wang et al 2017), or other categories. The “identifying
the subjects of opinions” part of the quest relates to the granularity of SA. Traditional SA
mostly focuses on document- or sentence-level sentiment and thus assumes a single subject
of opinions (Nazir et al 2022b; Liu et al 2020). In recent decades, the explosion of online
opinion text has attracted increasing interest in distilling more targeted insights on specific
entities or their aspects within each sentence through finer-grained SA (Nazir et al 2022b;
Liu et al 2020; Akhtar et al 2020; You et al 2022; Ettaleb et al 2022). This is the problem
ABSA aims to solve.

2.2 ABSA and its subtasks

ABSA involves identifying the sentiments toward specific entities or their attributes, called
aspects. These aspects can be explicitly mentioned in the text or implied from the context
(“implicit aspects”), and can be grouped into aspect categories (Nazir et al 2022a; Akhtar
et al 2020; Maitama et al 2020; Xu et al 2020b; Chauhan et al 2019; Akhtar et al 2018).
Appendix A.1 presents a more detailed definition of ABSA, including its key components
and examples.
A complete ABSA solution as described above traditionally involves a combination
of subtasks, with the fundamental ones (Li et al 2022a; Huan et al 2022; Li et al 2020;
Fei et al 2023b; Pathan and Prakash 2022) being Aspect (term) Extraction (AE), Opinion
(term) Extraction (OE), and Aspect-Sentiment Classification (ASC), or in an aggregated
form via Aspect-Category Detection (ACD) and Aspect Category Sentiment Analysis
(ACSA).
The choice of subtasks in an ABSA solution reflects both the problem formulation and,
to a large extent, the technologies and resources available at the time. The solutions to
these fundamental ABSA subtasks evolved from pure linguistic and statistical solutions to
296 Page 4 of 51 Y. C. Hua et al.

the dominant machine learning (ML) approaches (Maitama et al 2020; Cortis and Davis
2021; Liu et al 2020; Federici and Dragoni 2016), usually with multiple subtask models or
modules orchestrated in a pipeline (Li et al 2022b; Nazir and Rao 2022). More recently, the
rise of multi-task learning brought an increase in End-to-end (E2E) ABSA solutions that
can better capture the inter-task relations via shared learning (Liu et al 2024), and many
only involve a single model that provides the full ABSA solution via one composite task
(Huan et al 2022; Li et al 2022b; Zhang et al 2022b). The most typical composite ABSA
tasks include Aspect-Opinion Pair Extraction (AOPE) (Nazir and Rao 2022; Li et al 2022c;
Wu et al 2021), Aspect-Polarity Co-Extraction (APCE) (Huan et al 2022; He et al 2019),
Aspect-Sentiment Triplet Extraction (ASTE) (Huan et al 2022; Li et al 2022b; Du et al
2021; Fei et al 2023b), and Aspect-Sentiment Quadruplet Extraction/Prediction (ASQE/
ASQP) (Zhang et al 2022a; Lim and Buntine 2014; Zhang et al 2021a, 2024a). We provide
a more detailed introduction to ABSA subtasks in Appendix A.2.

2.3 The context‑ and domain‑dependency challenges

The nature and the interconnection of its components and subtasks determine that ABSA
is heavily domain- and context-dependent (Nazir et al 2022b; Chebolu et al 2023; Howard
et al 2022). Domain refers to the ABSA task (training or application) topic domains, and
context can be either the “global” context of the document or the “local” context from the
text surrounding a target word token or word chunks. At least in English, the same word
or phrase could mean different things or bear different sentiments depending on the con-
text and topic domains. For example, “a big fan” could be an electric appliance or a per-
son, depending on the sentence and the domain; “cold” could be positive for ice cream but
negative for customer service; and “DPS” (damage per second) could be either a gaming
aspect or non-aspect in other domains. Thus, the ability to incorporate relevant context is
essential for ABSA solutions; and those with zero or very small context windows, such as
n-gram and Markov models, are rare in ABSA literature and can only tackle a limited range
of subtasks (e.g. Presannakumar and Mohamed 2021).
Moreover, although many language models (e.g. Bidirectional Encoder Representations
from Transformers (BERT, Devlin et al 2019), Generative pre-trained transformers (GPT,
Brown et al 2020), recurrent neural network (RNN)-based models) already incorporated
local context from the input-sequence and/or general context through pre-trained embed-
dings, they still performed unsatisfactorily on some ABSA domains and subtasks, espe-
cially Implicit AE (IAE), AE with multi-word aspects, AE and ACD on mixed-domain
corpora, and context-dependent ASC (Phan and Ogunbona 2020; You et al 2022; Liang
et al 2022; Howard et al 2022). Many studies showed that ABSA task performance ben-
efits from expanding the feature space beyond the generic and input textual context. This
includes incorporating domain-specific dataset/representations and additional input fea-
tures such as Part-of-Speech (POS) tags, syntactic dependency relations, lexical databases,
and domain knowledge graphs or ontologies (Howard et al 2022; You et al 2022; Liang
et al 2022). Nonetheless, annotated datasets and domain-specific resources are costly to
produce and limited in availability, and domain adaptation, as one solution to this, has
been an ongoing challenge for ABSA (Chen and Qian 2022; Zhang et al 2022b; Nazir et al
2022b; Howard et al 2022; Satyarthi and Sharma 2023).
The above highlights the critical role of domain-specific datasets and resources in
ABSA solution quality, especially for supervised approaches. On the other hand, it suggests
the possibility that the prevalence of dataset-reliant solutions in the field, and a skewed
A systematic review of aspect‑based sentiment analysis: domains,… Page 5 of 51 296

ABSA dataset domain distribution, could systemically hinder ABSA solution performance
and generalisability (Chen and Qian 2022; Fei et al 2023a), thus confining ABSA research
and solutions close to the resource-rich domains and languages. This idea underpins this
literature review’s motivation and research questions.

2.4 Review rationale

This review is motivated by the following rationales:

First, the shift towards ML, especially supervised and/or DL solutions for ABSA, high-
lights the importance of dataset resources. In particular, annotated large benchmark data-
sets are crucial for the quality and development of ABSA research. Meanwhile, the finer
granularity of ABSA also brings the persistent challenge of domain dependency described
in Sect. 2.3. The diversity of ABSA datasets and their domains can have a direct and sys-
tematic impact on research and applications.
The early seminal works in ABSA were motivated by commercial applications and
focused on product and service reviews (Liu et al 2020; Rana and Cheah 2016; Do et al
2019), such as Ganu et al (2009), Hu and Liu (2004b), and Pontiki et al (2014, 2015, 2016)
that laid influential foundations with widely-used product and service review ABSA bench-
mark datasets (Rana and Cheah 2016; Do et al 2019). Nevertheless, the need for mining
insights from opinions far exceeds this single domain. Many other areas, especially the
public sector, also have an abundance of opinionated text data and can benefit from ABSA,
such as helping policy-makers understand public attitudes and reactions towards events or
changes (Sharma and Shekhar 2020), improving healthcare services and treatments via
patient experience and concerns in clinical visits, symptoms, drug efficacy and side-effects
(Cavalcanti and Prudêncio 2017; Gui and He 2021), and guiding educators in meeting
teacher and learner needs and improving their experience (Wankhade et al 2022; Tubishat
et al 2021; García-Pablos et al 2018; Poria et al 2016). While the more general SA research
has been applied to “nearly every domain” (Nazir et al 2022b, p. 1), this does not seem to
be the case for ABSA. Chebolu et al (2023) reviewed 62 public ABSA datasets released
between 2004 and 2020 covering “over 25 domains” (Chebolu et al 2023, p. 1). However,
53 out of these 62 datasets were reviews of restaurants, hotels, and digital products; only
five were not related to commercial products or services, and merely one was on the public
sector domain (university reviews).
The above-mentioned evidence raises questions: Will this dataset domain homogeneity
be found with a larger sample of primary studies? Does this domain skewness reflect the
concentration of ABSA research focus or merely the lack of dataset diversity? This moti-
vated our RQ1 (“To what extent is ABSA research and its dataset resources dominated by
the commercial (especially the product and service review) domain? What proportion of
ABSA research focuses on other domains and dataset resources?”) Answers to these ques-
tions could inform and shape future ABSA research through individual research decisions
and community resource collaboration.
Second, there are many good survey papers on ABSA, most focused on introducing
the methodological details of common ABSA subtasks and solutions (e.g. Maitama et al
2020; Sabeeh and Dewang 2019; Rana and Cheah 2016; Soni and Rambola 2022; Gangan-
war and Rajalakshmi 2019; Zhou et al 2019) or specific approaches such as DL methods
for ABSA (e.g. Liu et al 2020; Do et al 2019; Wang et al 2021a; Chen and Fnu 2022;
Mughal et al 2024; Zhang et al 2022c; Satyarthi and Sharma 2023). We list these sur-
veys in Appendix A.3 as additional resources for the reader. Nonetheless, many of these
296 Page 6 of 51 Y. C. Hua et al.

reviews only explored each subtask and/or technique individually and often by iterating
through reviewed studies, and few examined their combinations or changes over time and
with quantitative evidence. For example, although the above-listed reviews (Liu et al 2020;
Do et al 2019; Wang et al 2021a; Chen and Fnu 2022) reported the rise of DL approaches
in ABSA similar to that of NLP as a whole, it is unclear whether ABSA research was
also increasingly dominated by the attention mechanism from the Transformer architecture
(Vaswani et al 2017) and pre-trained large language models since 2018 (Manning 2022),
and if linguistic and traditional ML approaches were still active. In addition, most of these
surveys used a smaller and selected sample that could not support conclusions on trends.
As the field matures, we believe it is necessary and important to examine trends and mat-
ters outside the problem solution itself, so as to inform research decisions, identify issues,
and call for necessary community awareness and actions. We thus proposed RQ2 (“What
are the most common ABSA problem formulations via subtask combinations, and what pro-
portion of ABSA studies only focus on a specific sub-task?”) and RQ3 (“What is the trend
in the ABSA solution approaches over time? Are linguistic and traditional machine-learn-
ing approaches still in use?”).
In order to identify patterns and trends for our RQs, a sufficiently sized representative
sample and systematic approach are required. We chose to conduct an SLR, as this type of
review aims to answer specific research questions from all available primary research evi-
dence following well-defined review protocols (Kitchenham and Charters 2007). Moreo-
ver, none of the existing SLRs on ABSA share the same focus and RQs as ours: Among
the 192 survey/review papers obtained from four major digital database searches detailed
in Sect. 3, only eight were SLRs on ABSA, within which four focused on non-English
language(s) (Alyami et al 2022; Obiedat et al 2021; Hoti et al 2022; Rani and Kumar
2019), two on specific domains (software development, social media) (Cortis and Davis
2021; Lin et al 2022), one on a single subtask (Maitama et al 2020), and one mentioned
ABSA subtasks as a side-note under the main topic of SA (Ligthart et al 2021).
In summary, this review aims to address gaps in the ABSA literature. The high-level
nature of our research questions is best answered through a large-scale SLR to provide
solid evidence. The next section presents our SLR approach and sample.

3 Methods

Following the guidance of Kitchenham and Charters (2007), we conducted this SLR with
pre-planned scope, criteria, and procedures highlighted below. The complete SLR methods
and process are detailed in Appendix B.

3.1 Main procedures

For the main SLR sample, we sourced the primary studies in October 2022 from four
major peer-reviewed digital databases: ACM Digital Library, IEEE Xplore, Science Direct,
and SpringerLink. First, we manually searched and extracted 4191 database results without
publication-year constraints. Appendix B.1 provides more details of the search strategies
A systematic review of aspect‑based sentiment analysis: domains,… Page 7 of 51 296

Table 1  Inclusion and exclusion criteria used for this Systematic Literature Review (SLR)
Inclusion criteria Exclusion criteria

1. Published in the English language 1. The main text of the article is not in the English
language
2. Has both the sentiment analysis component and 2. Missing either the sentiment analysis or entity/
entity/aspect-level granularity aspect-level granularity in the research focus
3. Focuses on text data 3. Only contains search keyword in the reference
section
4. Is a primary study with quantitative elements in 4. Only contains less than 5 ABSA-related search
the ABSA approach keywords outside the reference section
5. Contains original solutions or ideas for ABSA 5. Is not a primary study (e.g. review articles, meta-
task(s) analysis)
6. Involves experiment and results on ABSA task(s) 6. Does not provide quantitative experiment results
on ABSA task(s)
7. Has fewer than three pages
8. Contains multimodal (i.e. non-text) input data for
ABSA task(s)
9. The research focus is not on ABSA task(s), even
though ABSA models might be involved (e.g.
recommender system)
10. Focuses on transferring existing ABSA
approaches between languages
11. The ABSA tasks are integrated into a model built
for other purposes, and there are no stand-alone
ABSA method details and/or evaluation results

and results. Next, we applied the inclusion and exclusion criteria listed in Table 1 via auto-
matic1 and manual screening steps and identified 519 in-scope peer-reviewed research pub-
lications for the review. The complete screening process, including that of the automatic
screening, is described in Appendix B.2. We then manually reviewed the in-scope primary
studies and recorded data following a planned scheme. Lastly, we checked, cleaned, and
processed the extracted data and performed quantitative analysis against our RQs.

3.2 Main SLR sample summary

Figure 1 shows the number of total reviewed vs. included studies across all publica-
tion years for the 4191 SLR search results. The search results include studies published
between 1995 and 2023 (N = 1), although all of the pre-2008 ones (2 from the 90s, 8 from
2003–2006, 17 from 2007) were not ABSA-focused and were excluded during automatic
screening. The earliest in-scope ABSA study in the sample was published in 2008, fol-
lowed by a very sparse period until 2013. The numbers of extracted and in-scope publi-
cations have both grown noticeably since 2014, a likely result of the emergence of deep
learning approaches, especially sequence models such as RNNs (Manning 2022; Sutskever
et al 2014). We also present a breakdown of the included studies by publication year and
type in Figure 9 in Appendix C.

1
Our PDF mining for automatic review screening code is available at https://​doi.​org/​10.​5281/​zenodo.​
12872​948.
296 Page 8 of 51 Y. C. Hua et al.

Fig. 1  Number of studies by publication year: total reviewed (N = 4191) vs. included (N = 519)

3.3 Note on “domain” mapping

In order to answer RQ1, we made the distinction between “research application domain”
(“research domain” in short) and “dataset domain”, and manually examined and classified
each study and its datasets into domain categories.
We considered each study’s research domain to be “non-specific” unless the study men-
tioned a specific application domain or use case as its motivation. For the dataset domain,
we examined each dataset used by our sample, standardised its name, and recorded the
domain from which it was drawn/selected based on the description provided by the author
or the dataset source webpage. Datasets without a specific domain (e.g. Twitter tweets
crawled without a specific domain filter) were labelled as “non-specific”.
We then manually grouped the research and dataset domains into 19 common catego-
ries used for analysis. More details and examples on domain mapping are available in
Appendix B.3.

3.4 Phase‑2 targeted review on in‑context learning

Additionally, generative “foundation models” (Bommasani et al 2022), defined as models

with billions of parameters pre-trained on enormous general-purpose data and adaptable
to diverse downstream NLP tasks, have become ubiquitous after our SLR data collection
(e.g. ChatGPT OpenAI 2023, released in November 2022). We use the term “foundation
models” to distinguish them from the earlier pre-trained Large Language Models (LLMs)
such as BERT (Devlin et al 2019), BART (Lewis et al 2020), and T5 (Raffel et al 2020),
which have relatively fewer parameters and typically require fine-tuning for task adaptation
(Zhang et al 2022c). These generative foundation models brought a new paradigm of “In-
context Learning” (ICL) (Brown et al 2020, p. 4), where task adaptation can occur solely
via conditioning the model on the text input instructions (“prompts”) with zero (“zero-shot
ICL”) or few (“few-shot ICL”) examples and no model parameter changes (Brown et al
A systematic review of aspect‑based sentiment analysis: domains,… Page 9 of 51 296

2020; Dong et al 2024). To capture and analyse this new development while balancing fea-
sibility and currency, we conducted a Phase-2 targeted review in July 2024.
This Phase-2 targeted review focuses solely on the ICL implementations of pre-trained
generative models for ABSA tasks, excluding those involving fine-tuning to draw a distinc-
tion from other non-ICL deep-learning approaches covered in the SLR. To extend the SLR
sample beyond the original extraction time, we conducted a new database search2 in July
2024 for studies published from 2022 onwards and removed the ones already included in
the SLR sample. The new search results were screened using the SLR criteria described in
Table 1 and then combined with the 519 SLR final samples. We then applied an additional
filtering condition “Gen-LLM” to all the in-scope ABSA primary studies, which further
selected publications with at least one occurrence of any of the following keywords outside
the Reference section: “generative”, “in-context”, “in context learning”, “genai”, “bart”,
“t5”, “flan-t5”, “gpt”, “chatgpt”, “llama”, and “mistral”. With the help of our automatic
screening pipeline detailed in Appendix B.2, we were able to efficiently auto-screen the
new search results and re-screen the previous SLR sample for the ”Gen-LLM” keywords in
less than one hour.
In total, the new search yielded 271 additional in-scope ABSA primary studies from
4359 search results. After applying the “Gen-LLM” filtering condition to the combined
790 in-scope ABSA primary studies, we obtained 208 Phase-2 samples for manual review,
which comprised 91 studies from the new search and 117 from the earlier SLR sample,
ranging from 2008 to 2024. The Phase-2 targeted review results are presented in Sect. 4.5.
Unless specified otherwise, the results below only refer to those of the SLR.

4 Results

This section presents the SLR results corresponding to each of the RQs:

4.1 Results for RQ1

RQ1. To what extent is ABSA research and its dataset resources dominated by the
commercial (especially the product and service review) domain? What proportion of
ABSA research focuses on other domains and dataset resources?
To answer RQ1, we examined the distribution of reviewed studies by their research (appli-
cation) domains, dataset domains, and the relationship between the two. From the 519
reviewed studies, we recorded 218 datasets, 19 domain categories (15 research domains
and 17 dataset domains), and obtained 1179 distinct “study-dataset” pairs and 630 unique
“study & dataset-domain” combinations. The key results are summarised below and pre-
sented in Table 2 and Fig. 2. We also list the datasets used by more than one reviewed
study in the Appendix Table 15.
In summary, our results answer RQ1 by showing that: (1) The majority (65.32%)
of the reviewed studies were not for any specific application domain and only 24.28%
targeted “product/service review”. (2) The dataset resources used in the sample were
mostly domain-specific (84.44%) and dominated by the “product/service review”

2
This new database search followed the same procedures and criteria as the SLR, except that we aborted
the SpringerLink search due to persistent database interface search result navigation issues during our data
collection period.
296 Page 10 of 51 Y. C. Hua et al.

Table 2  Number of in-scope studies per each research (application) and dataset domain category
Domain Count of studies % of studies per Count of stud- % of studies per
per research research domain ies per dataset dataset domain (%)
domain (%) domain

Non-specific 339 65.32 98 15.56

Product/service review 126 24.28 447 70.95
Student feedback/educa- 12 2.31 19 3.02
tion review
Politics/policy-reaction 8 1.54 5 0.79
Healthcare/medicine 7 1.35 9 1.43
Video/movie review 6 1.16 19 3.02
News 5 0.96 8 1.27
Finance 5 0.96 4 0.63
Research/academic 3 0.58 3 0.48
reviews
Disease 3 0.58 3 0.48
Employer review 1 0.19 1 0.16
Nuclear energy 1 0.19 0 0.00
Natural disaster 1 0.19 1 0.16
Music review 1 0.19 1 0.16
Multiple domain 1 0.19 0 0.00
Location review 0 0.00 8 1.27
Book review 0 0.00 2 0.32
Biology 0 0.00 1 0.16
Singer review 0 0.00 1 0.16
Total 519 100.00 630 100.00

Bold indicates the highest value in each column, as a way to highlight the mismatch between research and
dataset domains

datasets (70.95%). (3) Both the research effort and dataset resources were scant in the
non-commercial domains, especially the main public sector areas, with fewer than 13
studies across 14 years in each of the healthcare, policy, and education domains, where
about half of the used datasets were created from scratch for the study.
Beyond RQ1, (1) and (2) above also suggest a significant mismatch between the
research and dataset domains as visualised in Fig. 2. Further, when filtering out data-
sets used by less than 10 studies, we discovered an alarming lack of dataset diversity as
only 12 datasets remained, of which 10 were product/service reviews. When examining
the three-way relationship among research domain, dataset domain, and dataset name,
we further identified an over-representation (78.20%) of the four SemEval restaurant
and laptop review benchmark datasets. This is illustrated in Fig. 4.

4.1.1 Detailed results for RQ1

For research (application) domains indicated by the stated research use case or motivation,
the majority (65.32%, N = 339) of the 519 reviewed studies have a “non-specific” research
A systematic review of aspect‑based sentiment analysis: domains,… Page 11 of 51 296

Fig. 2  Distribution of unique “study–dataset” pairs (N = 1179, with 519 studies and 218 datasets) by
research (application) domains (left) and dataset domains (right). Note (1) The top flow visualises a mis-
match between the two domains: the majority of studies without a specific research domain used datasets
from the product/service review domain. (2) The disproportionately small number of samples in both
domains that were neither “non-specific” nor “product/service review”

Fig. 3  Number of in-scope studies by research (application) domain and publication year ( N = 518). This
graph excludes the one 2023 study (extracted in October 2022) to avoid trend confusion

domain, followed by just a quarter (24.28%, N = 126) in the “product/service review” cat-
egory. However, the number of studies in the rest of the research domains is magnitudes
smaller in comparison, with only 12 studies (2.31%) in the third largest category “student
feedback/education review” since 2008, followed by 8 in Politics/policy-reaction (1.54%),
and only 7 in Healthcare/medicine (1.35%). Figure 3 revealed further insights from the
296 Page 12 of 51 Y. C. Hua et al.

trend of research domain categories with five or more reviewed studies. Interestingly,
“product/service review” has been a persistently major category over time, and has only
been consistently taken over by “non-specific” since 2015. The sharp increase of domain-
“non-specific” studies since 2018 could be partly driven by the rise of pre-trained language
models such as BERT and the greater sequence processing power from the Transformer
architecture and the attention mechanism (Manning 2022), as more researchers explore the
technicalities of ABSA solutions.
As to the dataset domains, Table 2 suggests that among the 630 unique “study & data-
set-domain” pairs, the majority (70.95%, N = 447) are in the “product/service review” cat-
egory, followed by 15.56% (N = 98) in “Non-specific”. The third place is shared by two
magnitude-smaller categories: “student feedback/ education review” (3.02%, N = 19) and
“video/movie review” (3.02%, N = 19). The numbers of studies with datasets from the
Healthcare/medicine (1.43%, N = 9) and Politics/policy-reaction (0.79%, N = 5) domains
were again single-digit. Moreover, nearly half of the unique datasets in the public domains
were created by the authors for the first time: 5/9 in Healthcare/medicine, 2/4 in Politics/
policy-reaction, and 8/12 in Student feedback/ Education review.
Furthermore, to understand the dataset diversity across samples and domains, we
grouped the 1179 unique “study-dataset” pairs by “research-domain, dataset-domain, data-
set-name” combinations and zoomed into the 757 entries with ten or more study counts
each. As shown in Table 3 and illustrated in Fig. 4, among these 757 unique combinations,
95.77% ( N = 725) are in the “non-specific” research domain, of which 90.48% (N = 656)
used “product/service review” datasets. Most interestingly, these 757 entries only involve
12 distinct datasets of which 10 were product and service reviews, and 78.20% (N = 592)
are taken up by the four SemEval datasets from the early pioneer work (Pontiki et al 2014,
2015, 2016) mentioned in Sect. 2.4: SemEval 2014 Restaurant, SemEval 2014 Laptop
(these two alone account for 50.33% of all 757 entries), SemEval 2016 Restaurant, and
SemEval 2015 Restaurant. This finding echos (Xing et al 2020; Chebolu et al 2023): “The
SemEval challenge datasets... are the most extensively used corpora for aspect-based senti-
ment analysis” (Chebolu et al 2023, p.4). Meanwhile, the top dataset used under “product/
service review” research and dataset domains is the original product review dataset created
by the researchers. Chebolu et al (2023) and Wikipedia (2023) provides a detailed intro-
duction to the SemEval datasets.
It is noteworthy that among the 519 reviewed studies, 20 focused on cross-domain
or domain-agnostic ABSA, and 19 of them did not have a specific research application
domain. However, while all 20 studies used multiple datasets, 17 solely involved the “prod-
uct/service review” domain category by using reviews of restaurants and different prod-
ucts, and 14 used at least one SemEval dataset. The only three studies that went beyond
the “product/service review” dataset domain added in movie reviews, singer reviews, and
generic tweets.

4.2 Results for RQ2

RQ2. What are the most common ABSA problem formulations via subtask combina-
tions, and what proportion of ABSA studies only focus on a specific sub-task?
For RQ2, we examined the 13 recorded subtasks and 805 unique “study-subtask” pairs
to identify the most explored ABSA subtasks and subtask combinations across the 519
reviewed studies. As shown in Fig. 5a, 32.37% ( N = 168) of the studies developed
Table 3  Number of studies per each research (application) domain, dataset domain, and dataset combination for all datasets used by ten or more in-scope studies (N = 757)
Research domain Dataset domain Dataset Count of studies % of studies (%)

Non-specific Product/service review SemEval 2014 Restaurant 200 26.42

Non-specific Product/service review SemEval 2014 Laptop 181 23.91
Non-specific Product/service review SemEval 2016 Restaurant 110 14.53
non-specific product/service review SemEval 2015 Restaurant 101 13.34
Non-specific Non-specific Twitter (Dong et al. 2014) 53 7.00
Non-specific Product/service review Amazon customer review datasets (Hu and Liu 25 3.30
2004a)
Product/service review Product/service review Product review (original) 17 2.25
A systematic review of aspect‑based sentiment analysis: domains,…

Non-specific Non-specific Twitter (original) 16 2.11

Non-specific Product/service review SemEval 2015 Laptop 16 2.11
Product/service review Product/service review Amazon product review (original) 15 1.98
Non-specific Product/service review Yelp Dataset Challenge Reviews 12 1.59
Non-specific Product/service review SemEval 2016 Laptop 11 1.45
Total 757 100.00
Page 13 of 51 296
296 Page 14 of 51 Y. C. Hua et al.

Fig. 4  Number of studies per each research (application) domain (left), dataset domain (middle), and data-
set (right) combination, filtered by datasets used by 10 or more in-scope studies ( N = 757). The three-way
relationship highlights that not only did the majority of the sample studies with “non-specific” research
domain use datasets from the ‘product/service review‘ domain, but their datasets were also dominated by
only four SemEval datasets on two types of product and service reviews

Fig. 5  Number of studies by ABSA subtask

A systematic review of aspect‑based sentiment analysis: domains,… Page 15 of 51 296

Fig. 6  Distribution of unique “Study–ABSA subtask” pairs by publication year ( N = 805). This graph
excludes the one 2023 study (extracted in October 2022) to avoid trend confusion

full-ABSA solutions through the combination of AE and ASC, and a similar proportion
(30.83%, N = 160) focused on ASC alone, usually formulating the research problem as
contextualised sentiment analysis with given aspects and the full input text. Only 15.22%
(N = 79) of the studies solely explored the AE problem. This is consistent with the number
of studies by individual subtasks shown in Fig. 5b, where ASC is the most explored sub-
task, followed by AE and ACD.
Moreover, Fig. 6 reveals a small but noticeable rise in composite subtask ASTE since
2020 (N = 1, 5 and 10 in 2017, 2021, 2022) and a decline in ASC and AE around the
same period. This could signify a problem formulation shift driven by deep-learning, espe-
cially multi-task learning methods for E2E ABSA. Our Phase-2 targeted review findings in
Sect. 4.5 add more insights into this.

4.3 Results for RQ3

RQ3. What is the trend in the ABSA solution approaches over time? Are linguistic
and traditional machine-learning approaches still in use?
To answer RQ3, we examined the 519 in-scope studies along two dimensions, which we
call “paradigm” and “approach”. We use “paradigm” to indicate whether a study employed
techniques along the supervised-unsupervised dimension and other types, such as rein-
forcement learning. We classify non-machine-learning approaches under the “unsuper-
vised” paradigm, as our focus is on dataset and resource dependency. By “approach”,
we refer to the more specific type of techniques, such as deep learning (DL), traditional
machine learning (traditional ML), linguistic rules (“rules” for short), syntactic features
and relations (“syntactics” for short), lexicon lists or databases (“lexicon” for short), and
ontology or knowledge-driven approaches (“ontology” for short).
Overall, the results suggest that our samples are dominated by fully- (60.89%) and
partially-supervised (5.40%) ML methods that are more reliant on annotated datasets and
prone to their impact. As to ABSA solution approaches, the sample shows that DL methods
296 Page 16 of 51 Y. C. Hua et al.

Table 4  Number of studies by Paradigm Count of studies % of studies (%)

paradigm (N = 519)
Fully-supervised 316 60.89
Unsupervised 97 18.69
Hybrid 73 14.07
Semi-supervised 22 4.24
Weakly-supervised 6 1.16
Reinforcement learning 3 0.58
Self-supervised 2 0.39
Total 519 100.00

The unsupervised category includes non-ML approaches

Fig. 7  Number of studies using DL and traditional ML approaches

have rapidly overtaken traditional ML methods since 2017, particularly with the preva-
lent RNN family (55.91%) and its combination with the fast-surging attention mechanism
(26.52%). Meanwhile, traditional ML and linguistic approaches have remained a small but
steady force even in the most recent years. Context engineering through introducing lin-
guistic and knowledge features to DL and traditional ML approaches was very common.
More detailed results and richer findings are presented below.

4.3.1 Paradigms

Table 4 lists the number of studies per each of the main paradigms. Among the 519
reviewed studies, 66.28% ( N = 344) is taken up by those using somewhat- (i.e. fully-,
semi- and weakly-) supervised paradigms that have varied levels of dependency on labelled
datasets, where the fully-supervised ones alone account for 60.89% (N = 316). Only
A systematic review of aspect‑based sentiment analysis: domains,… Page 17 of 51 296

Table 5  Number of studies by paradigm and approaches (N = 519)

Paradigm Approaches Count of studies % of studies (%)

Supervised DL 149 47.15

Syntactics, DL 53 16.77
DL, traditional ML 23 7.28
Other (< 5 each) 91 28.80
Total 316 100.00
Semi-supervised DL 7 31.82
Syntactics, Lexicon, traditional ML 3 13.64
traditional ML 2 9.09
Rules, Syntactics, Lexicon 2 9.09
Other (< 5 each) 8 36.36
Total 22 100.00
Weakly-supervised DL, traditional ML 2 33.33
Traditional ML 1 16.67
DL 1 16.67
Syntactics, DL 1 16.67
Syntactics, Ontology, DL 1 16.67
Other (< 5 each) 0 0.00
Total 6 100.00
Self-supervised DL 2 100.00
Other (< 5 each) 0 0.00
Total 2 100.00
Unsupervised Rules, Syntactics, Lexicon 24 24.74
Rules, Syntactics, Lexicon, Ontology 15 15.46
Rules, Syntactics 8 8.25
Traditional ML 7 7.22
Other (< 5 each) 43 44.33
Total 97 100.00
Hybrid Rules, Syntactics, traditional ML 6 8.22
Rules, Syntactics, Lexicon, traditional ML 6 8.22
Traditional ML 5 6.85
Syntactics, traditional ML 4 5.48
Rules, Syntactics, Lexicon 4 5.48
Rules, Syntactics, Lexicon, Ontology, 4 5.48
traditional ML
Other (< 5 each) 44 60.27
Total 73 100.00
Reinforcement learning DL 1 33.33
Syntactics, DL 1 33.33
Other (< 5 each) 1 33.33
Total 3 100.00
TOTAL 519 100.00

Bold indicates subtotals of the corresponding rows

296 Page 18 of 51 Y. C. Hua et al.

19.65% (N = 102) of the studies do not require labelled data, which are mostly unsuper-
vised (18.69%, N = 97). In addition, hybrid studies are the third largest group (14.07%,
N = 73).
We further analysed the approaches under each paradigm and focused on three for more
details: deep learning (DL), traditional machine learning (ML), and Linguistic and Statisti-
cal Approaches. The results are detailed below and presented in Fig. 7 and Tables 5, 6.

4.4 Approaches

As shown in Fig. 7a and Table 5, among the 519 reviewed studies, 60.31% (N = 313)
employed DL approaches, and 30.83% ( N = 160) are DL-only. The DL-only approach is
particularly prominent among fully-supervised (47.15%, N = 149) and semi-supervised
(31.82%, N = 7) studies. Supplementing DL with syntactical features is also the second
most popular approach in fully-supervised studies (16.77%, N = 53).

(1) DL Approaches

Figure 7a suggests that the 313 studies involving DL approaches are dominated by Recur-
rent Neural Network (RNN)-based solutions (55.91%, N = 175), of which nearly half
used a combination of RNN and the attention mechanism (26.52%, N = 83), followed by
attention-only (19.17%, N = 60) and RNN-only (9.90%, N = 31) models. The RNN family
mainly consists of Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM),
and Gated Recurrent Unit (GRU). These neural-networks are featured by sequential pro-
cessing that captures temporal dependencies of text tokens, and can thus incorporate sur-
rounding text as context for prediction (Liu et al 2020; Satyarthi and Sharma 2023). On the
other hand, the sequential nature poses challenges with parallelisation and the exploding
and vanishing gradient problems associated with long sequences (Vaswani et al 2017; Liu
et al 2020). Although LSTM and GRU can mitigate these issues somewhat through cell
state and memory controls, efficiency and long-dependency challenges still hinder their
performance (Vaswani et al 2017; Liu et al 2020; Satyarthi and Sharma 2023). The atten-
tion mechanism complements RNNs by dynamically updating weights across the input
sequence based on each element’s relevance to the current task, and thus guides the model
to focus on the most relevant elements (Vaswani et al 2017).
In addition, convolutional and graph-neural approaches [e.g. convolutional neural net-
works (CNN), graph neural networks (GNN), graph convolutional networks (GCN)] also
play smaller but noticeable roles in DL-based ABSA studies. While CNN was commonly
used as an alternative to the sequence models such as RNNs (Liu et al 2020; J et al 2021;
Zhang et al 2020), the graph-based networks (GNN, GCN) were mainly used to model the
non-linear relationships such as external conceptual knowledge (e.g. Liang et al 2022) and
syntactic dependency structures (e.g. Fei et al 2023a, b; Li et al 2022c) that are not well
captured by the sequential networks like RNNs and the flat structure of the attention mod-
ules. As a result, they inject richer context into the overall learning process (Du et al 2021;
Xu et al 2020a; Wang et al 2022).
Figure 8 depicts the trend of the main approaches across the publication years. We
excluded the one study pre-published for 2023 to avoid confusing trends. It is clear that
DL approaches have risen sharply and taken dominance since 2017, mainly driven by the
rapid growth in RNN- and attention-based studies. This coincides with the appearance of
the Transformer architecture in 2017 (Vaswani et al 2017) and the resulting pre-trained
A systematic review of aspect‑based sentiment analysis: domains,… Page 19 of 51 296

Fig. 8  Distribution of studies per method by publication year (N = 1017 with 519 unique studies). This
graph excludes the one 2023 study (extracted in October 2022) to avoid trend confusion

models such as BERT (Devlin et al 2019) that were a popular embedding choice to be used
alongside RNNs in DL and hybrid approaches (e.g., Li et al 2021; Zhang et al 2022a).
GNN/GCN-based approaches remain small in number but have noticeable growth since
2020 (N = 2, 2, 16, 24 in each of 2019–2022, respectively), suggesting an increased
effort to dynamically integrate relational context into the learning process within the DL
framework.

(2) Traditional ML approaches

Interestingly, as shown in Fig. 8 traditional ML approaches remain a steady force over the
decades despite the rapid rise of DL methods. Table 5 and Fig. 7b provide some insight
into this: Among the 519 reviewed studies, while 60.31% employed DL approaches as
mentioned in the previous sub-section, over half (54.53%, N = 283) also included tra-
ditional ML approaches, with the top 3 being Support Vector Machine (SVM; 20.14%,
N = 57), Conditional Random Field (CRF; 14.49%, N = 41), and Latent Dirichlet allo-
cation (LDA; 12.72%, N = 36). Table 5 suggests that among the major paradigms, tradi-
tional ML were often used in combination with DL approaches for fully-supervised studies
(7.28%, N = 23), and along with linguistic rules, syntactic features, and/or lexicons and
ontology in hybrid studies (27.40%, N = 20). Across all paradigms, traditional ML-only
approaches are relatively rare (max N = 7).

(3) Linguistic and statistical approaches

While Table 5 illustrates the prevalence of fusing ML approaches with linguistic and sta-
tistical features or modules, there were 67 studies (12.91% out of the total 519) on pure
296 Page 20 of 51 Y. C. Hua et al.

Table 6  Number of studies by pure linguistic or statistical approaches and publication year (N = 67)
Approach 2011–2013 2014–2016 2017–2019 2020–2022 Total Total%

Syntactics, Rules, Lexicon 1 8 8 6 23 34.33%

Syntactics, Rules, Lexicon, Ontol- 1 3 9 3 16 23.88%
ogy
Syntactics, Rules 2 4 5 11 16.42%
Statistical 2 1 1 4 5.97%
Syntactics, Rules, Lexicon, Statisti- 1 1 1 3 4.48%
cal
Syntactics, Rules, Statistical 2 1 3 4.48%
Lexicon 1 1 1.49%
Rules 1 1 1.49%
Rules, Lexicon 1 1 1.49%
Rules, Lexicon, Statistical 1 1 1.49%
Syntactics, Lexicon 1 1 1.49%
Syntactics, Lexicon, Statistical 1 1 1.49%
Syntactics, Statistical 1 1 1.49%
TOTAL 6 16 26 19 67 100.00%

linguistic or statistical approaches. As shown in Table 6, although small in number, these

non-ML approaches have persisted over time. The most popular combination (34.33%,
N = 23) was rules built on syntactic features (e.g. POS tags and dependency parse trees)
and used along lexicon resources (e.g. domain-specific aspect lists, SentiWordNet,3
MPQA4). A typical example is using POS tags and/or lexicon resources to narrow the
scope of aspect or opinion term candidates, applying further rules based on POS tags or
dependency relations for AE or OE, and/or using lexicon resources for candidate pruning,
categorisation, or sentiment labelling (e.g. Asghar et al 2019; Dragoni et al 2019; Nawaz
et al 2020). The second top combination (23.88%, N = 16) is the above-mentioned one
plus ontology (e.g. domain-specific ontology, ConceptNet,5 WordNet6) to bring in exter-
nal knowledge of concepts and relations (e.g. Federici and Dragoni 2016; Marstawi et al
2017). Pure statistical methods were relatively rare (5.97%, N = 4), and mainly included
frequency-based methods such as N-gram and TF-IDF, and other statistical modelling
methods that were not commonly seen in the ML field.

4.5 ICL and generative approach in ABSA ‑ Phase‑2 targeted review results

ICL is a subgenre of the DL approach. However, we discuss the relevant results in this
separate subsection due to the Phase-2 review’s more focused sample and finer granular-
ity. Despite the trending popularity of the ICL approach in NLP research and applications
since 2022 (Dong et al 2024), our results suggest that the ABSA research community is
just beginning to explore it with caution. Among the 208 ABSA studies from 2008 to 2024

3
https://​github.​com/​aesuli/​Senti​WordN​et.
4
https://​mpqa.​cs.​pitt.​edu/.
5
https://​conce​ptnet.​io/.
6
https://​wordn​et.​princ​eton.​edu/.
Table 7  Studies with in-context learning (ICL) approach on ABSA tasks (N = 5 out of 208 samples from 2008–2024)
Paper Task Non-ICL models ICL models ICL approach Result

Zhou et al (2024) ASQE RoBERTa ChatGPT 5-shot ICL ChatGPT performed worse than almost all
non-ICL methods with about 20–30%
lower micro-F1
Liu et al (2024) ASTE BERT ChatGPT Zero-shot, 5-shot ICL ChatGPT 5-shot ICL performed better than
0-shot ICL, but was still around 20% lower
in F1 score than the main method
Su et al (2024) ASC T5 Llama2-7b-chat, Llama2- N.A ChatGPT 3.5 performed slightly better than
13b-chat, ChatGPT-3.5 the Llama-2 models across datasets, but
was still up to about 20% lower than the
main model in accuracy and F1
Amin et al (2024) AE, ASC, OE LSTM, RoBERTa (not fine-tuned) GPT 3.5-Turbo, GPT-4 Zero-shot On AE and ASC tasks, RoBERTa per-
formed significantly better than the two
GPT models with up to about 20% higher
accuracy. GPT 3.5-Turbo performed better
A systematic review of aspect‑based sentiment analysis: domains,…

than GPT-4 on AE and ASC tasks, and

outperformed all the other models with the
OE task with up to 15% higher accuracy
Mughal et al (2024) ASC, ACSA LSTM, Flan-T5, DeBERTa, PaLM GPT 3.5-Turbo, PaLM-bison Zero-shot PaLM showed the best overall performance
on ASC and ASCA tasks in terms of accu-
racy and F1, closely followed by fine-tuned
DeBERTa. However, PaLM had the lowest
accuracy and F1 scores with the multi-
aspect multi-sentiment (MAMS) dataset,
with 21–48% lower accuracy and F1 than
fine-tuned DeBERTa

The Non-ICL Models were trained or fine-tuned unless specified otherwise

Page 21 of 51 296
296 Page 22 of 51 Y. C. Hua et al.

Table 8  Studies with fine-tuned generative large language models (LLMs) on ABSA tasks (N = 18 out of
208 samples from 2008–2024)
Paper Task GenAI model

Hoang et al (2022) ASQE BART​

Kang et al (2022) OE BART​
Gong and Li (2022) ASTE BERT
William and Khodra (2022) ASTE T5
Lil et al (2023) ASQE BART (with BERT for upstream
embedding)
Li et al (2023) ASQE BART​
­ ABSAa
Suchrady and Purwarianti (2023) ASTE, AOPE, AE, OE, U mT5b
Yu et al (2023a) AESCc, AOPE, ASTE BART + GAT (Graph Attention
Networks)
Yan et al (2023) ASC,ASTE T5
Yu et al (2023b) AE, OE, ASQE T5 (data generator + self-training) +
BERT pair-classifier as descrimi-
nator
Lee and Kim (2023) ASTE, ASQE, ASC T5
Liu et al (2024) ASTE BERT + GPN (Global Pointer
Network)
Lee et al (2024) AE, ASC GPT-2 fine-tuned by LoRA
Dang et al (2024) AE, OE, ASC, ACD, AOPE, mT5, ­ViT5b
ACSA, ASTE, ACSD, ASQE
Su et al (2024) ASC T5
Zhou et al (2024) ASQE RoBERTa
Wang et al (2024) AE, OE, ACD, ASC, ACSA, T5
­ SPEd,
AOPE, ASTE, ASQE, A
­CSPEe
Zhang et al (2024b) ASQE T5
a
UABSA Unified aspect-based sentiment analysis
b
mT5, ViT5: multi-lingual variants of T5
c
AESC aspect term extraction and sentiment classification
d
ASPE aspect sentiment pair extraction
e
CSPE category sentiment pair extraction

containing at least one occurrence of the “Gen-LLM” keywords, only five (all published in
2024) applied ICL to both composite and traditional ABSA tasks. All of these studies were
exploring the performance of foundation models via ICL against other approaches, rather
than focusing on an ICL ABSA solution. Table 7 summarises the models, ABSA tasks,
and key findings from these studies. Overall, four of the five studies found that zero-shot
and even 5-shot ICL on foundation models (mainly GPTs) could not reach the performance
of fine-tuned or fully trained DL models, especially those leveraging pre-trained LLMs to
fine-tune a contextual-embedding.
In addition, we identified an emerging trend by examining the Phase-2 review non-ICL
samples: Those employing fine-tuned generative LLMs mostly formulated the ABSA tasks
as Sequence-to-Sequence (Seq2Seq) text generation problems, with a particular focus on
composite tasks such as ASTE and ASQE. As shown in Table 8, within the 208 samples,
a total of 18 studies (all from the new search) published in 2022–2024 applied pre-trained
A systematic review of aspect‑based sentiment analysis: domains,… Page 23 of 51 296

generative LLMs with fine-tuning. The majority of these studies used models based on
T5 (N = 9) and BART (N = 5) with the full Transformer (Vaswani et al 2017) encoder-
decoder architecture, followed by encoder-only (N = 3, BERT and RoBERTa Liu et al
2019) and decoder-only (N = 1, GPT-2 Radford et al 2019) models. All but two of these 18
studies were on composite ABSA tasks, mainly ASTE and ASQE. Moreover, two studies
(Yu et al 2023b; Zhang et al 2024b) also leveraged the generative capability of these LLMs
to augment training data to enrich the fine-tuned embedding.
Compared with this Seq2Seq generation approach, the common applications of pre-
trained LLMs in earlier studies from the main SLR sample often formulate the ABSA task
as a classification problem (Zhang et al 2022c). These studies mostly use encoder-only
LLMs for their pre-trained representations to fine-tune a contextual embedding (Zhang
et al 2022c), which is then connected to other context-injection or relationship-learning
modules and a classifier output layer. For instance, Zhang et al (2022a) employed pre-
trained BERT with BiLSTM, a feed-forward neural network (FFNN), and CRF. Li et al
(2021) used pre-trained BERT as an encoder and a decoder featuring a GRU. In contrast,
the Seq2Seq generative approach can be illustrated by the signature “Generative Aspect-
based Sentiment analysis (GAS)” proposed by Zhang et al (2021b), which leveraged the
LLM’s pre-trained and fine-tuned encoder module for context-aware embedding and used
the fine-tuned decoder module to generate text representations of the label sets (e.g., tri-
plets) or as annotations next to the original input text (Zhang et al 2021b, 2022c).

5 Discussion

This review was motivated by the literature gap in capturing trends in ABSA research to
answer higher-level questions beyond technical details, and the concern that the domain-
dependent nature could predispose ABSA research to systemic hindrance from a combina-
tion of resource-reliant approaches and skewed resource domain distribution. By system-
atically reviewing the two waves of 727 in-scope primary studies published between 2008
and 2024, our quantitative analysis results identified trends in ABSA solution approaches,
confirmed the above-mentioned concern, and provided detailed insights into the relevant
issues. In this section, we examine the primary findings, share ideas for future research,
and reflect on the limitations of this review.

5.1 Significant findings and trends

5.1.1 The out‑of‑sync research and dataset domains

Under RQ1, we examined the distributions of and relationships between our sample’s
research (application) domains and dataset domains. The results showed strong skewness
in both types of domains and a significant mismatch between them: While the majority
(65.32%, N = 339) of the 519 studies did not aim for a specific research domain, a greater
proportion (70.95%, N = 447) used datasets from the “product/service review” domain. A
closer inspection of the link between the two domains revealed a clear mismatch: Among
the 757 unique “research-domain, dataset-domain, dataset-name” combinations with ten
or more studies: 90.48% ( N = 656) of the studies in the “non-specific” research domain
(95.77%, N = 725) used datasets from the “product/service review” domain. This suggests
296 Page 24 of 51 Y. C. Hua et al.

that the lack of non-commercial-domain datasets could have forced generic technical stud-
ies to use benchmark datasets from a single popular domain. Given ABSA problem’s
domain-dependent nature, this could have indirectly hindered the solution development and
evaluation across domains.
The results also showed that the other important and prevalent ABSA application
domains such as education, medicine/healthcare, and public policy, were clearly under-
researched and under-resourced. Among the reviewed samples from these three public-sec-
tor domains, about half of their datasets were created for the studies by their authors, indi-
cating a lack of public dataset resources, hence the cost and challenge of developing ABSA
research in these areas. As a likely consequence, even the most researched domain among
these three had only 12 studies (2.31% out of 519) since 2008. The dataset resource scar-
city in these public sector domains deserves more research community attention and sup-
port, especially given these domains’ overall low research resources vs. the high cost and
domain knowledge required for quality data annotation. In particular, for domains such as
“Student feedback/education review” that often face strict data privacy and consent restric-
tions, it is crucial that the ABSA research community focus on creating ethical and open-
access datasets to leverage community resources.

5.1.2 The dominance and limitations of the SemEval datasets

The results under RQ1 also revealed further issues with dataset diversity, even within the
dominant “product/service review” domain. Out of the 757 unique “research-domain, data-
set-domain, dataset-name” combinations with ten or more studies, 78.20% ( N = 592) are
taken up by the four popular SemEval datasets: The SemEval 2014 Restaurant and Laptop
datasets alone account for 50.33% of all 757 entries, and the other two (SemEval 2015 and
2016 Restaurant).
The level of dominance of the SemEval datasets is alerting, not only because of their
narrow domain range, but also for the inheritance and impact of the SemEval datasets’
limitations. Several studies (e.g. Chebolu et al 2023; Xing et al 2020; Jiang et al 2019;
Fei et al 2023a) suggest that these datasets fail to capture sufficient complexity and granu-
larity of the real-world ABSA scenarios, as they primarily only include single-aspect or
multi-aspect-but-same-polarity sentences, and thus mainly reflect sentence-level ABSA
tasks and ignored subtasks such as multi-aspect multi-sentiment ABSA. The experiment
results from Xing et al (2020), Jiang et al (2019) and Fei et al (2023a) consistently showed
that all 35 ABSA models (including those that were state-of-the-art at the time) (9 in Xing
et al 2020, 16 in Jiang et al 2019, 10 in Fei et al 2023a) that were trained and performed
well on the SemEval 2014 ABSA datasets showed various extents of performance drop (by
up to 69.73% in Xing et al 2020) when tested on same-source datasets created for more
complex ABSA subtasks and robustness challenges. Given that the SemEval datasets are
heavily used as both training data and “benchmark” to measure ABSA solution perfor-
mance, their limitations and prevalence are likely to form a self-reinforcing loop that con-
fines ABSA research. To break free from this dataset-performance self-reinforcing loop, it
is critical that the ABSA research community be aware of this issue, and develop and adopt
datasets and practices that are robustness-oriented, such as the automatic data-generation
framework and the resulting Aspect Robustness Test Set (ARTS) developed by Xing et al
(2020) for probing model robustness in distinguishing target and non-target aspects, and
the Multi-Aspect Multi-Sentiment (MAMS) dataset created by Jiang et al (2019) to reflect
A systematic review of aspect‑based sentiment analysis: domains,… Page 25 of 51 296

more realistic challenges and complexities in aspect-term sentiment analysis (ATSA) and
aspect-category sentiment analysis (ACSA) tasks.

5.1.3 The reliance on labelled‑datasets

The domain and dataset issues discussed above would not be as problematic if most ABSA
studies employed methods that are dataset-agnostic. However, our results under RQ3 show the
opposite. Only 19.65% (N = 102, with 97 being unsupervised) of the 519 reviewed studies
do not require labelled data, whereas 66.28% (N = 344) are somewhat-supervised, and fully-
supervised studies alone account for 60.89% (N = 316).
As demonstrated in Sect. 2.3, the domain can directly affect whether a chunk of text is
considered an aspect or the relevant sentiment term, and plays a crucial role in contextual
inferences such as implicit aspect extraction and multi-aspect multi-sentiment pairing. The
domain knowledge reflected via ABSA labelled datasets can further shape the linguistic rules,
lexicons, and knowledge graphs for non-machine-learning approaches; and define the under-
pinning feature space, representations, and acquired relationships and inferences for trained
machine-learning models. When applying a solution built on datasets from a domain that is
very remote from or much narrower than the intended application domain, it is predictable that
the solution performance would be capped at subpar and even fail at more context-heavy tasks
(Phan and Ogunbona 2020; You et al 2022; Liang et al 2022; Howard et al 2022; Chen and
Qian 2022; Zhang et al 2022b; Nazir et al 2022b). Thus, domain transfer is crucially necessary
for balancing the uneven ABSA research and resource distributions across domains. However,
our finding that 17 out of the 20 reviewed cross-domain or domain-agnostic ABSA studies
solely used datasets from the “product/service review” domain raised questions about these
approaches’ generalisability and robustness in other domains, as well as whether such dataset
choices became another reinforcement of concentrating research effort and benchmarks within
this one dominant domain.
The rapid rise of deep learning (DL) in ABSA research could further add to the challenge
of overcoming the negative impact of this domain mismatch and dataset limitations via the
non-linear multi-layer dissemination of bias in the representation and learned relations, thus
making problem-tracking and solution-targeting difficult. In reality, of the 519 reviewed stud-
ies, 60.31% (N = 313) employed DL approaches, and nearly half (47.15%, N = 149) of the
fully-supervised studies and 30.83% (N = 160) of all reviewed studies were DL-only.
Moreover, RNN-based solutions dominate the DL approaches (55.91%, N = 175), mainly
with the RNN-attention combination (26.52%, N = 83) and RNN-only (9.90%, N = 31) mod-
els. RNN and its variants such as LSTM, BiLSTM, and GRU are known for their limitations
in capturing long-distance relations due to their sequential nature and the subsequent memory
constraints (Vaswani et al 2017; Liu et al 2020). Although the addition of the attention mecha-
nism enhances the model’s focus on more important features such as aspect terms (Vaswani
et al 2017; Liu et al 2020), traditional attention weights calculation struggles with multi-word
aspects or multi-aspect sentences (Liu et al 2020; Fan et al 2018). In addition, whilst 16.77%
( N = 53) of the fully-supervised studies introduced syntactical features to their DL solutions,
additional features also increased the input size. According to Prather et al (2020), sequen-
tial models, even the state-of-the-art LLMs, showed impaired performance as the input grew
longer and could not always benefit from additional features.
296 Page 26 of 51 Y. C. Hua et al.

5.1.4 The potential of generative LLMs and foundation models

Lastly, the Phase-2 targeted review highlights the ABSA community’s caution toward the
direct adoption of generative foundation models, with only five out of 208 recent stud-
ies testing the ICL approach and most yielding subpar results compared to other methods.
However, most of these studies only tested zero-shot instructions with simple model set-
tings. It is worth further exploring the potential of foundational models and ICL in ABSA
by focusing more on instruction and example engineering, model parameter optimisation,
and task re-formulation (Dong et al 2024).
On the other hand, the fine-tuning of smaller generative LLMs has seen increasing
adoption through the “ABSA as Seq2Seq text generation” approach, demonstrating promis-
ing task performance. Although this generative approach can incorporate data augmenta-
tion and self-training to reduce reliance on labelled datasets, the cost of fine-tuning, the
need for labelled base data, and the domain-transfer problem remain significant challenges
(Zhang et al 2022c). In this context, the task adaptability and multi-domain pre-trained
knowledge of foundation models could provide potential solutions.
As Zhang et al (2022c) noted, progress in applying pre-trained LLMs and foundation
models to ABSA could be impeded by dataset resources constraints. To match the param-
eter size of these models, more diverse, complex, and larger datasets are required for effec-
tive fine-tuning or comprehensive testing. In low-resource domains where dataset resources
are already limited, this requirement could further complicate the adoption of these tech-
nologies (Satyarthi and Sharma 2023).

5.2 Ideas for future research

Overall, by adopting a “systematic perspective, i.e., model, data, and training” (Fei et al
2023a, p.28) combined with a quantitative approach, we identified high-level trends
unveiling the development and direction of ABSA research, and found clear evidence
of large-scale issues that affect the majority of the existing ABSA research. The skewed
domain distributions of resources and benchmarks could also restrict the choice of new
studies. On the other hand, this evidence also highlights areas that need more attention
and exploration, including: ABSA solutions and resource development for the less-stud-
ied domains (e.g. education and public health), low-resource and/or data-agnostic ABSA,
domain adaptation, alternative training schemes such as adversarial (e.g. Fei et al 2023a;
Chen et al 2021) and reinforcement learning (e.g. Vasanthi et al 2022; Wang et al 2021b),
and more effective feature and knowledge injection. Future research could contribute to
addressing these issues by focusing on ethically producing and sharing more diverse and
challenging datasets in minority domains such as education and public health, improving
data synthesis and augmentation techniques, exploring methods that are less data-depend-
ent and resource-intensive, and leveraging the rapid advancements in pre-trained LLMs
and foundation models.
In addition, our results also revealed emerging trends and new ideas. The relatively
recent growth of end-to-end models and composite ABSA subtasks provide opportunities
for further exploration and evaluation. The fact that hybrid approaches with non-machine-
learning techniques and non-textual features remain steady forces in the field after nearly
three decades suggests valuable characteristics that are worth re-examining under the light
of new paradigms and techniques. Moreover, the small number of Phase-2 samples using
A systematic review of aspect‑based sentiment analysis: domains,… Page 27 of 51 296

ICL and fine-tuning generative LLM approaches may suggest that we have only captured
early adopters. More thorough exploration of these approaches and continued tracking of
their development alongside other methods are necessary to understand how the ABSA
community can leverage the resources and capabilities embedded within LLMs and foun-
dation models.
Lastly, it is crucial that the community invest in solution robustness, especially for
machine-learning approaches (Xing et al 2020; Jiang et al 2019; Fei et al 2023a). This
could mean critical examination of the choice of evaluation metrics, tasks, and bench-
marks, and being conscious of their limitations vs. the real-world challenges. The “State-
Of-The-Art” (SOTA) performance based on certain benchmark datasets should never
become the motivation and holy grail of research, especially in fields like ABSA where the
real use cases are often complex and even SOTA models do not generalise far beyond the
training datasets. More attention and effort should be paid to analysing the limitations and
mistakes of ABSA solutions, and drawing from the ideas of other disciplines and areas to
fill the gaps.

5.3 Limitations

We acknowledge the following limitations of this review: First, our sample scope is by
no means exhaustive, as it only includes primary studies from four peer-reviewed dig-
ital databases and only those published in the English language. Although this can be
representative of a core proportion of ABSA research, it does not generalise beyond
this without assumptions. The “peer-reviewed” criteria also meant that we overlooked
preprint servers such as arXiv.org that more closely track the latest development of
ML and NLP research. Second, no search string is perfect. Our database search syntax
and auto-screening keywords represent our best effort in capturing ABSA primary
studies, but may have missed some relevant ones, especially with the artificial “total
pages < 3” and “total keyword (except SA, OM) outside Reference < 5” exclusion
criteria. Moreover, our search completeness might have been affected by the perfor-
mance of the database search engines. This is evidenced by the significant number of
extracted search results that were entirely irrelevant to the search keywords, as well as
our abandonment of the 2024 SpringerLink search due to interface issues. Enhance-
ments in digital database search capabilities could significantly improve the effective-
ness and reliability of future literature review studies, particularly SLRs. Third, we
may have missed datasets, paradigms, and approaches that are not clearly described in
the primary studies, and our categorisation of them is also subject to the limitations
of our knowledge and decisions. Future review studies could consider a more innova-
tive approach to enhance analytical precision and efficiency, such as applying ABSA
and text summarisation alongside the screening and reviewing process. Fourth, we
did not compare solution performance across studies due to the review focus, sam-
ple size, and the variability in experimental settings across studies. Evaluating the
effectiveness of comparable methods and the suitability of evaluation metrics would
enhance our findings and offer more valuable insights.
296 Page 28 of 51 Y. C. Hua et al.

6 Conclusion

ABSA research is riding the wave of the explosion of online digital opinionated text data
and the rapid development of NLP resources and ideas. However, its context- and domain-
dependent nature and the complexity and inter-relations among its subtasks pose chal-
lenges to improving ABSA solutions and applying them to a wider range of domains. In
this review, we systematically examined existing ABSA literature in terms of their research
application domain, dataset domain, and research methodologies. The results suggest a
number of potential systemic issues in the ABSA research literature, including the pre-
dominance of the “product/service review” dataset domain among the majority of studies
that did not have a specific research application domain, coupled with the prevalence of
dataset-reliant methods such as supervised machine learning. We discussed the implication
of these issues to ABSA research and applications, as well as their implicit effect in shap-
ing the future of this research field through the mutual reinforcement between resources
and methodologies. We suggested areas that need future research attention and proposed
ideas for exploration.

Appendix A: Aspect‑based sentiment analysis (ABSA)

Appendix A.1: Definition and examples

Aspect-based sentiment analysis (ABSA) is a sub-domain of fine-grained SA (Nazir et al

2022a). ABSA focuses on identifying the sentiments towards specific entities or their
attributes/ features called aspects (Nazir et al 2022a; Akhtar et al 2020). An aspect can be
explicitly expressed in the text (explicit aspect) or absent from the text but implied from
the context (implicit aspects) (Maitama et al 2020; Xu et al 2020b). Moreover, the aspect-
level sentiment could differ across aspects and be different from the overall sentiment of
the sentence or the document (e.g. Akhtar et al 2020, 2017; Li et al 2022a). Some studies
further distinguish aspect into aspect term and aspect category, with the former referring
to the aspect expression in the input text (e.g. “pizza”), and the latter a latent construct that
is usually a high-level category across aspect terms (e.g. “food”) that are either identified
or given (Chauhan et al 2019; Akhtar et al 2018).
The following examples illustrate the ABSA terminologies:

Example 1 (From a restaurant review7): “The restaurant was expensive, but the menu was
great.” This sentence has one explicit aspect “menu” (sentiment term: “great”, sentiment
polarity: positive), one implicit aspect “price” (sentiment term: “expensive”, sentiment
polarity: negative). Depending on the target/given categories, the aspects can be further
classified into categories, such as “menu” into “general” and “price” into “price”.

Example 2 (From a laptop review8): “It is extremely portable and easily connects to WIFI
at the library and elsewhere.” This sentence has two implicit aspects: “portability” (sen-
timent term: “portable”, sentiment polarity: positive), “connectivity” (sentiment term:

7
https://​alt.​qcri.​org/​semev​al2014/​task4/.
8
https://​alt.​qcri.​org/​semev​al2015/​task12/.
A systematic review of aspect‑based sentiment analysis: domains,… Page 29 of 51 296

“easily”, sentiment polarity: positive). The aspects can be further classified into categories,
such as both under “laptop” (as opposed to “software” or “support”).

Example 3 (Text from a course review): “It was too difficult and had an insane amount of
work, I wouldn’t recommend it to new students even though the tutorial and the lecturer
were really helpful.” The two explicit aspects in Example 3 are “tutorial” and “lecturer”
(sentiment terms: “helpful”, polarities: positive). The implicit aspects are “content” (sen-
timent term: “too difficult”, sentiment polarity: negative), “workload” (sentiment term:
“insane amount”, sentiment polarity: negative), and “course” (sentiment term: “would not
recommend”, sentiment polarity: negative). An illustration of aspect categories would be
assigning the aspect “lecturer” to the more general category “staff” and “tutorial” to the
category “course component”.

As demonstrated above, the fine granularity makes ABSA more targetable and informa-
tive than document- or sentence-level SA. Thus, ABSA can precede downstream applica-
tions such as attribute weighting in overall review ratings (e.g. Da’u et al 2020), aspect-
based opinion summarisation (e.g. Yauris and Khodra 2017; Kumar et al 2022; Almatrafi
and Johri 2022), and automated personalised recommendation systems (e.g. Ma et al 2017;
Nawaz et al 2020).
Compared with document- or sentence-level SA, while being the most detailed and
informative, ABSA is also the most complex and challenging (Huan et al 2022). The most
noticeable challenges include the number of ABSA subtasks, their interrelations and con-
text dependencies, and the generalisability of solutions across topic domains.

Appendix A.2: ABSA Subtasks

A full ABSA solution has more subtasks than coarser-grained SA. The most fundamental
ones (Li et al 2022a; Huan et al 2022; Li et al 2020; Fei et al 2023b) include:

Aspect (term) extraction/identification (AE), which has a slight variation in meaning

depending on the overall ABSA approach. Some authors (e.g. (Zhang et al 2023; Luo
et al 2019; Ruskanda et al 2019)) consider AE as identifying the attribute or entity that
is the target of an opinion expressed in the text and sometimes call it “opinion target
extraction” (Guo et al 2018). In these cases, opinion terms were often identified in order
to find their target aspect terms. Others (e.g. Akhtar et al 2020; Gunes 2016; Li et al
2020; Ettaleb et al 2022; Tran et al 2020) define AE as identifying the key or all attrib-
utes of entities mentioned in the text. Implicit-Aspect Extraction (IAE) is often men-
tioned as a task by itself due to its technical challenge.
Opinion (term) Extraction/Identification (OE), which relates to identifying the “opin-
ion terms” or the sentiment expression of a specific entity/aspect (e.g. Li et al 2022a;
Wang et al 2018; Fernando et al 2019; Fei et al 2023b). In Example 1 above, an OE task
would extract the sentiment terms “great” (associated with the aspect term “menu”) and
“expensive” (associated with the implicit aspect “price”).
Aspect-Sentiment Classification (ASC), which refers to obtaining the sentiment polarity
category (e.g. negative, neutral, positive, conflict) or sentiment score (e.g. 1 to 5 or −1
to 1 along the scale from negative to positive) associated with a given aspect or aspect
category (e.g. Akhtar et al 2020; Gojali and Khodra 2016; Castellanos et al 2011).
This is often done via evaluating the associated opinion term(s), and sentiment lexicon
296 Page 30 of 51 Y. C. Hua et al.

resources such as the SentiWordNet (Baccianella et al 2010) and SenticNet (Cambria

et al 2016) can be used to assign polarity scores (Gojali and Khodra 2016). Sentiment
scores can be further aggregated across opinion terms for the same aspect, or across
aspect terms to generate higher-level ratings, such as aspect-category ratings within or
across documents (Gojali and Khodra 2016; Castellanos et al 2011).
As an extension of AE, some studies also involve Aspect-Category Detection (ACD) and
Aspect Category Sentiment Analysis (ACSA) when the focus of sentiment analysis is on
(often pre-defined) latent topics or concepts and requires classifying aspect terms into
categories (Pathan and Prakash 2022).

Traditional full ABSA solutions often perform the subtasks in a pipeline manner (Li et al
2022b; Nazir and Rao 2022) using one or more of the linguistic (e.g. lexicons, syntactic
rules, dependency relations), statistical (e.g. n-gram, Hidden Markov Model (HMM)), and
machine-learning approaches (Maitama et al 2020; Cortis and Davis 2021; Federici and
Dragoni 2016). For instance, for AE and OE, some studies used linguistic rules and senti-
ment lexicons to first identify opinion terms and then the associated aspect terms of each
opinion term, or vice versa (e.g. You et al 2022; Cavalcanti and Prudêncio 2017), and then
moved on to ASC or ACD using a supervised model or unsupervised clustering and/or
ontology (Nawaz et al 2020; Gojali and Khodra 2016). Hybrid approaches are common
given the task combinations in a pipeline.
With the rise of multi-task learning and deep learning (Chen et al 2022), an increas-
ing number of studies explore ABSA under an End-to-end (E2E) framework that performs
multiple fundamental ABSA subtasks in one model to better capture the inter-task relations
(Liu et al 2024), and some combine them into a single composite task (Huan et al 2022; Li
et al 2022b; Zhang et al 2022b). These composite tasks are most commonly formulated as
a sequence- or span-based tagging problem (Huan et al 2022; Li et al 2022b; Nazir and Rao
2022). The most common composite tasks are: Aspect-Opinion Pair Extraction (AOPE),
which directly outputs {aspect, opinion} pairs from text input (Nazir and Rao 2022; Li
et al 2022c; Wu et al 2021) such as “⟨menu, great⟩” from Example 1; Aspect-Polarity Co-
Extraction (APCE) (Huan et al 2022; He et al 2019), which outputs {aspect, sentiment
polarity} pairs such as “⟨menu, positive⟩”; Aspect-Sentiment Triplet Extraction (ASTE)
(Huan et al 2022; Li et al 2022b; Du et al 2021; Fei et al 2023b), which outputs {aspect,
opinion, sentiment category} triplets, such as “⟨menu, great, positive⟩”; and Aspect-Senti-
ment Quadruplet Extraction/Prediction (ASQE/ASQP) (Zhang et al 2022a; Lim and Bun-
tine 2014; Zhang et al 2021a, 2024a) that outputs {aspect, opinion, aspect category, senti-
ment category} quadruplets, such as “⟨menu, great, general, positive⟩”.

Appendix A.3: Other ABSA reviews

As this review focuses on trends instead of detailed solutions and methodologies, we refer
interested readers to existing review papers that provide comprehensive and in-depth sum-
maries of common ABSA subtask solutions and approaches, for example:

• Explicit and implicit AE: Rana and Cheah (2016), Ganganwar and Rajalakshmi (2019),
Soni and Rambola (2022), Maitama et al (2020).
• Deep learning (DL) methods for ABSA: Do et al (2019), Liu et al (2020), Wang et al
(2021a), Chen and Fnu (2022), Zhang et al (2022c), Mughal et al (2024). Specifically:
Table 9  Digital databases and search details used for this systematic literature review (SLR)
Database Search string Search criteria PDFs exported Type of exported publications

ACM Digital Library ((“aspect based” OR “aspect-based”) AND (senti- ∙ No year filter 1514 ∙ 201 articles
ment OR extraction OR extract OR mining)) ∙ Search scope = Full article ∙ 1283 conference papers
OR “opinion mining” ∙ Content Type = Research Article ∙ 30 newsletters
∙ Media Format = PDF
∙ Publications = Journals OR Proceedings OR
Newsletters
IEEE Xplore ((“aspect based” OR “aspect-based”) AND (senti- ∙ Year filter = 2004–2022 (pilot search suggested 1639 ∙ 165 articles
ment OR extraction OR extract OR mining)) that 1995–2003 results were irrelevant) ∙ 1445 conference papers
OR “opinion mining” ∙ Search scope = Full article ∙ 29 magazine pieces
∙ Publications = Not “Books”
A systematic review of aspect‑based sentiment analysis: domains,…

Science Direct (“aspect based” OR “aspect-based”) AND (senti- ∙ No year filter 497 ∙ 497 articles
ment OR extraction OR extract OR mining)) ∙ Search scope = Title, abstract or author-speci-
OR “opinion mining” fied keywords
∙ Publications = Research Articles only
SpringerLink (“aspect based” OR “aspect-based”) AND (senti- ∙ No year filter 541 ∙ 218 articles
ment OR extraction OR extract OR mining)) ∙ English results only ∙ 323 conference papers
OR “opinion mining” ∙ Publications = Article, Conference Paper
Page 31 of 51 296
296 Page 32 of 51 Y. C. Hua et al.

– DL methods for ASC: Zhou et al (2019), Satyarthi and Sharma (2023).

– E2E ABSA, composite tasks, and pre-trained Large Language Models (LLMs) in
ABSA: Zhang et al (2022c) provided a comprehensive review and shared extensive
reading lists and dataset resource links via https://​github.​com/​IsakZ​hang/​ABSA-​
Survey. Mughal et al (2024) introduced common benchmark datasets, including
more challenging ones for composite ABSA tasks. They also reviewed and tested
the ABSA task performance of representative RNN-based models and pre-trained
LLMs.
• Multimodal ABSA: Zhao et al (2024).

Appendix B: Full SLR methodology

This section provides a complete, detailed description of the SLR methodology and
procedures.

Appendix B.1: Research identification

To obtain the files for review, we conducted database searches between 24–25 October
2022, when we manually queried and exported a total of 4191 research papers’ PDF and
BibTeX (or the equivalent) files via the web interfaces of four databases. Table 9 details the
search string, search criteria, and the PDF files exported from each database.
Given the limited search parameters allowed in these digital databases, we adopted a
“search broad and filter later” strategy. These database search strings were selected based
on pilot trials to capture the ABSA topic name, the relatively prevalent yet unique ABSA
subtask term (“extraction”), and the interchangeable use between ABSA and opinion min-
ing; while avoiding generating false positives from the highly active, broader field of SA.
The “filter later” step was carried out during the “selection of primary studies” stage intro-
duced in the next section, which aimed at excluding cases where the keywords are only
mentioned in the reference list or sparsely mentioned as a side context, and opinion mining
studies that were at document or sentence levels.

Appendix B.2: Selection of primary studies

After obtaining the 4191 initial search results, we conducted a pilot manual file examina-
tion of 100 files to refine the pre-defined inclusion and exclusion criteria. We found that
some search results only contained the search keywords in the reference list or Appendix,
which was also reported in Prather et al (2020). In addition, there are a number of papers
that only mentioned ABSA-specific keywords in their literature review or introduction sec-
tions, and the studies themselves were on coarser-grained sentiment analysis or opinion
mining. Lastly, there were instances of very short research reports that provided insuffi-
cient details of the primary studies. Informed by these observations, we refined our inclu-
sion and exclusion criteria to those in Table 1 in Sect. 3. Note that we did not include
popularity criteria such as citation numbers so we can better identify novel practices and
A systematic review of aspect‑based sentiment analysis: domains,… Page 33 of 51 296

Table 10  The automatic and manual screening processes

Inclusion/exclusion types File count

Total extracted files 4191

Auto-excluded 3277
Step_1—Duplicate DOI across databases 19
Step_2—Duplicate Title across databases 11
Step_3—Survey/Not primary study papers 149
Step_4—Total keyword outside reference match = 0 26
Step_5—keyword matched only to SA and/or OM outside ­referencea 2235
Step_6—No sentiment∖W+analysis in keyword outside reference 243
Step_7—Total pages < 3 7
Step_8—Total keyword (except SA, OM) outside reference < 5a,b 587
Manually excluded 395
Type—Article withdrawn 1
Type—Not published in English 1
Type—Dataset unclear 1
Type—Low quality, unclear definition of ABSA 1
Type—No original method 2
Type—Duplicate of another included paper 3
Type—Sentence level SA 4
Type—Not text-data focused 6
Type—Not primary study 15
Type—Lack of details on ABSA tasks 31
Type—Review (Not primary study) 43
Type—No ABSA task experiment details/results 49
Type—Not ABSA focused; No ABSA task experiment details/results 238
Final included for review 519
a
SA, OM: the Regex keyword patterns ‘sentiment∖W+analysis’, ‘opinion∖W+mining’ respectively
b
The occurrence threshold 5 was chosen based on pilot file examination, which suggested that files with
target keyword occurrence below this threshold tended to be non-ABSA-focused
Bold indicates subtotals of the corresponding rows

avoid mainstream method over-dominance introduced by the citation chain (Chu and Evans
2021).
To implement the inclusion and exclusion criteria, we first applied PDF mining to auto-
matically exclude files that meet the exclusion criteria, and then refined the selection with
manual screening under the exclusion and inclusion criteria. Both of these processes are
detailed below. Our PDF mining for automatic review screening code is also available at
https://​doi.​org/​10.​5281/​zenodo.​12872​948.
The automatic screening consists of a pipeline with two Python packages: Pan-
das (Team 2023) and PyMuPDF.9 We first used Pandas to extract into a dataframe (i.e.
table) all exported papers’ file locations and key BibTex or equivalent information includ-
ing title, year, page number, DOI, and ISBN. Next, we used PyMuPDF to iterate through

9
https://​pypi.​org/​proje​ct/​PyMuP​DF/.
296 Page 34 of 51 Y. C. Hua et al.

each PDF file and add to the dataframe multiple data fields: whether the file was success-
fully decoded10 for text extraction (if marked unsuccessful, the file was marked for manual
screening), the occurrence count of each Regex keyword pattern listed below, and whether
each keyword occurs after the section headings that fit into Regex patterns that represent
variations of “references” and “bibliography” (referred to as “non-target sections” below).
We then marked the files for exclusion by evaluating the eight criteria listed under “Auto-
excluded” in Table 10 against the information recorded in the dataframe. Each of the auto-
exclusion results from Steps 1–4 and 7 in Table 10 were manually checked, and those
under Steps 5, 6, and 8 were spot-checked. These steps excluded 3277 out of the 4194
exported files.
Below are the regex patterns used for automatic keyword extraction and occurrence
calculation:
PDF search keyword Regex list: [’absa’, ’aspect∖W+base∖w*’, ’aspect∖W+extrac∖
w*’, ’aspect∖W+term∖w*’, ’aspect∖W+level∖w*’, ’term∖W+level’, ’sentiment∖
W+analysis’, ’opinion∖W+mining’]
For the 914 files filtered through the auto-exclusion process, we manually screened them
individually according to the inclusion and exclusion criteria. As shown in the second half
of Table 10, this final screening step refined the review scope to 519 papers.

Appendix B.3: Data extraction and synthesis

In the final step of the SLR, we manually reviewed each of the 519 in-scope publica-
tions and recorded information according to a pre-designed data extraction form. The key
information recorded includes each study’s research focus, research application domain
(“research domain” below), ABSA subtasks involved, name or description of all the data-
sets directly used, model name (for machine-learning solutions), architecture, whether a
certain approach or paradigm is present in the study (e.g. supervised learning, deep learn-
ing, end-to-end framework, ontology, rule-based, syntactic-components), and the specific
approach used (e.g. attention mechanism, Naïve Bayes classifier) under the deep learning
and traditional machine learning categories.
After the data extraction, we performed data cleaning to identify and fix recording
errors and inconsistencies, such as data entry typos and naming variations of the same
dataset across studies. Then we created two mappings for the research and dataset domains
described below.
For each reviewed study, its research domain was defaulted to “non-specific” unless the
study mentioned a specific application domain or use case as its motivation, in which case
that domain description was recorded instead.
The dataset domain was recorded and processed at the individual dataset level, as many
reviewed studies used multiple datasets. We standardised the recorded dataset names,
checked and verified the recorded dataset domain descriptions provided by the authors
or the source web-pages, and then manually categorised each domain description into a
domain category. For published/well-known datasets, we unified the recorded naming vari-
ations and checked the original datasets or their descriptions to verify the domain descrip-
tions. For datasets created (e.g. web-crawled) by the authors of the reviewed studies, we

10
https://​pdfmi​nersix.​readt​hedocs.​io/​en/​latest/​faq.​html.
A systematic review of aspect‑based sentiment analysis: domains,… Page 35 of 51 296

named them following the “[source] [domain] (original)” format, e.g. “Yelp restaurant
review (original) ”, or “Twitter (original)” if there was no distinct domain, and did not dif-
ferentiate among the same-name variations. In all of the above cases, if a dataset was not
created with a specific domain filter (e.g. general Twitter tweets), then it was classified as
“non-specific”.
The recorded research and dataset domain descriptions were then manually grouped into
19 common domain categories. We tried to maintain consistency between the research and
dataset domain categories. The following are two examples of possible mapping outcomes:

1. A study on a full ABSA solution without mentioning a specific application domain and
using Yelp restaurant review and Amazon product review datasets would be assigned a
research domain of “non-specific” and a dataset domain of “product/service review”.
2. A study mentioning “helping companies improve product design based on customer
reviews” as the motivation would have a research domain of “product/service review”,
and if they used a product review dataset and Twitter tweets crawled without filtering,
the dataset domains would be “product/service review” and “non-specific”.

After applying the above-mentioned standardisation and mappings, we analysed the syn-
thesised data quantitatively using the Pandas (Team 2023) library to obtain an overview of
the reviewed studies and explore the answers to our RQs.

Appendix C: Additional results

See Figs. 9, 10 and Tables 11, 12, 13, 14, 15.

Fig. 9  Number of included studies by publication year and type ( N = 519). Note Although our original
search scope included journal articles, conference papers, newsletters, and magazine articles, the final 519
in-scope studies consist of only journal articles and conference papers. Conference papers noticeably out-
numbered journal articles in all years until 2022, with the gap closing since 2016. We think this trend could
be due to multiple factors, such as the fact that our search was conducted in late October 2022 when some
conference publications were still not available; the publication lag for journal articles due to a longer pro-
cessing period; and potentially a change in publication channels that is outside the scope of this review
296 Page 36 of 51 Y. C. Hua et al.

Fig. 10  Number of included studies with the top 5 dataset languages by publication year
A systematic review of aspect‑based sentiment analysis: domains,… Page 37 of 51 296

Table 11  Number of studies by ABSA subtask combinations (N = 519)

Subtask combination Count of studies % of studies (%)

AE, ASC 168 32.37

ASC 160 30.83
AE 79 15.22
ACD, ASC 16 3.08
ASTE 15 2.89
AE, ACD, ASC 13 2.50
AE, OE, ASC 9 1.73
AE, ACD 9 1.73
AOPE 9 1.73
ACD 7 1.35
AE, OE 7 1.35
AE, OE, ACD 4 0.77
AE, OE, ACD, ASC 4 0.77
ASQE 2 0.39
AOPE, ASC 2 0.39
OE 1 0.19
OTE (opinion target extraction), ASC 1 0.19
Aspect-based embedding, ACD, ASC 1 0.19
ASC, OE 1 0.19
Aspect and synthetic sample discrimination 1 0.19
AE, OE, ASC, AOPE, ASTE 1 0.19
AOPE, ASC, ACD 1 0.19
AOPE, ACD 1 0.19
AE, review-level SA 1 0.19
ACD, AE, ASC 1 0.19
AE, ASC, Aspect-based sentence segmentation 1 0.19
AE, ASC, Aspect-based embedding 1 0.19
AE, ASC, ACD 1 0.19
AE, ACD, OE 1 0.19
Data augmentation 1 0.19
Total 519 100.00
a
This table corresponds to Fig. 5a
296 Page 38 of 51 Y. C. Hua et al.

Table 12  Number of studies by individual ABSA subtasks (N = 805)

Individual subtask Count of studies % of Studies (%)

ASC 381 47.33

AE 300 37.27
ACD 59 7.33
OE 28 3.48
ASTE 16 1.99
AOPE 14 1.74
ASQE 2 0.25
Aspect and synthetic sample discrimination 1 0.12
OTE (opinion target extraction) 1 0.12
Aspect-based sentence segmentation 1 0.12
Data augmentation 1 0.12
Review-level SA 1 0.12
Total 805 100
a
This table corresponds to Fig. 5b

Table 13  Number of studies by DL approach Count of studies % of studies (%)

deep learning (DL) approaches
(N = 313)
RNN, Attention 83 26.52
Attention 60 19.17
RNN 31 9.90
RNN, CNN, Attention 25 7.99
CNN 20 6.39
RNN, CNN 19 6.07
GNN/GCN, Attention 17 5.43
CNN, Attention 17 5.43
RNN, GNN/GCN, Attention 17 5.43
Other (< 5 % each) 24 7.67
Total 313 100.00
a
This table corresponds to Fig. 7a
A systematic review of aspect‑based sentiment analysis: domains,… Page 39 of 51 296

Table 14  Number of studies by traditional machine learning approaches (N = 283)

Traditional ML Count of studies % of studies (%)

Support vector machine (SVM) 57 20.14

Conditional random field (CRF) 41 14.49
Latent Dirichlet allocation (LDA) 36 12.72
Naïve Bayes (NB) 32 11.31
Random forest (RF) 24 8.48
Decision tree (DT) 18 6.36
Logistic regression (LR) 15 5.30
K-Nearest neighbors (KNN) 12 4.24
Multinomial Naïve Bayes (MNB) 10 3.53
K-means 5 1.77
Boosting 5 1.77
Other (N < 5) 28 9.89
Total 283 100.00
a
This table corresponds to Fig. 7b
296 Page 40 of 51 Y. C. Hua et al.

Table 15  Number of studies per dataset (N = 1179, with 519 studies and 218 datasets)
Datasets Count of studies % of studies (% )

SemEval 2014 Restaurant 211 17.90

SemEval 2014 Laptop 189 16.03
SemEval 2016 Restaurant 118 10.01
SemEval 2015 Restaurant 106 8.99
Twitter (Dong et al. 2014) 55 4.66
Amazon customer review datasets (Hu and Liu 2004a) 33 2.80
Amazon product review (original) 23 1.95
Product review (original) 22 1.87
Twitter (original) 21 1.78
SemEval 2015 Laptop 20 1.70
Yelp Dataset Challenge Reviews 16 1.36
SemEval 2016 Laptop 14 1.19
TripAdvisor Hotel review (original) 11 0.93
Hotel review (original) 10 0.85
TripAdvisor Hotel review (Wang et al. 2010) 9 0.76
TripAdvisor Restaurant review (original) 9 0.76
Movie review (original) 8 0.68
MAMS Multi-Aspect Multi-Sentiment dataset (Jiang et al. 2019) 8 0.68
SemEval 2016 Hotel 7 0.59
Twitter (Mitchell et al. 2013) 7 0.59
Amazon product review (McAuley et al. 2015) 7 0.59
VLSP 2018 Restaurant review 6 0.51
Restaurant review (Ganu et al. 2009) 6 0.51
Student feedback (original) 6 0.51
VLSP 2018 Hotel review 6 0.51
Chinese product review (Peng et al. 2018) 5 0.42
Stanford Twitter Sentiment / Sentiment140 (Go et al. 2009) 5 0.42
Yelp restaurant review (original) 5 0.42
SentiHood (Saeidi et al. 2016) 4 0.34
Game review (original) 3 0.25
Sanders Twitter Corpus (STC) (Sanders 2011) 3 0.25
TripAdvisor Tourist review (original) 3 0.25
Coursera course review (original) 3 0.25
Online drug review (original) 3 0.25
Restaurant review (original) 3 0.25
Chinese Restaurant review (original) 3 0.25
Financial Tweets and News Headlines dataset (FiQA 2018) 2 0.17
SentiRuEval-2015 2 0.17
Amazon laptop reviews (He and McAuley 2016) 2 0.17
Hindi ABSA dataset (Akhtar et al. 2016) 2 0.17
Hotel review (Kaggle) 2 0.17
Service review (Toprak etc. 2010) 2 0.17
Indonesian Product review (original) 2 0.17
Indonesian Tweets—Political topic (original) 2 0.17
A systematic review of aspect‑based sentiment analysis: domains,… Page 41 of 51 296

Table 15  (continued)
Datasets Count of studies % of studies (% )

Steam game review (original) 2 0.17

Stanford Sentiment Treebank review data (Socher et al. 2013) 2 0.17
Kaggle movie review 2 0.17
Zomato restaurant review (original) 2 0.17
Turkish Product review (original) 2 0.17
SemEval 2015 Hotel 2 0.17
Weibo comments (original) 2 0.17
Amazon 50-domain Product review (Chen and Liu 2014) 2 0.17
Bangla ABSA Cricket, Restaurant dataset (Rahman and Dey 2018) 2 0.17
YouTube comments (original) 2 0.17
Beer review (McAuley et al. 2012) 2 0.17
CCF BDCI 2018 Chinese auto review dataset 2 0.17
ReLi Portuguese book reviews (Freitas et al. 2014) 2 0.17
Twitter product review (original) 2 0.17
SemEval 2016 Tweets 2 0.17
Product review dataset (Liu et al. 2015) 2 0.17
Product review dataset (Cruz et al. 2014) 2 0.17
Chinese Product review (original) 2 0.17
ICLR Open Reviews dataset 2 0.17
Amazon Electronics review data (Jo and Oh 2011) 2 0.17
Amazon product review (Wang et al. 2011) 2 0.17
SemEval 2015 (unspecified) 2 0.17
Product & service review (Kaggle) 2 0.17
Othe datasets (N = 1 each) 149 12.64
Total 1179 100.00

Author contributions Y. C. H designed, conducted, and wrote this review. P. D., J. W., and K. T. guided
the design of the review methodology and review protocol, and reviewed and provided feedback on the
manuscript.

Funding Open Access funding enabled and organized by CAUL and its Member Institutions.

Declarations
Conflict of interest The authors have no Conflict of interest as defined by Springer, or other interests that
might be perceived to influence the results and/or discussion reported in this paper.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License,
which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long
as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Com-
mons licence, and indicate if changes were made. The images or other third party material in this article
are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly
from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
296 Page 42 of 51 Y. C. Hua et al.

References
Akhtar MS, Ekbal A, Bhattacharyya P (2016) Aspect based sentiment analysis in Hindi: Resource creation
and evaluation. In: Calzolari N, Choukri K, Declerck T, et al (eds) Proceedings of the Tenth Interna-
tional Conference on Language Resources and Evaluation (LREC’16). European Language Resources
Association (ELRA), Portoroˇz, Slovenia, pp 2703–2709. https://​aclan​tholo​gy.​org/​L16-​1429
Akhtar MS, Gupta D, Ekbal A et al (2017) Feature selection and ensemble construction: a two-step method
for aspect based sentiment analysis. Knowl Based Syst 125:116–135. https://​doi.​org/​10.​1016/j.​kno-
sys.​2017.​03.​020. https://​linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S0950​70511​73014​8X
Akhtar MS, Ekbal A, Bhattacharyya P (2018) Aspect based sentiment analysis: category detection and sentiment
classification for Hindi. In: Gelbukh A (ed) Computational linguistics and intelligent text processing, vol
9624. Lecture Notes in Computer Science. Springer, Cham, pp 246–257. https://​doi.​org/​10.​1007/​978-3-​319-​
75487-1_​19
Akhtar MS, Garg T, Ekbal A (2020) Multi-task learning for aspect term extraction and aspect sentiment
classification. Neurocomputing 398:247–256. https://​doi.​org/​10.​1016/j.​neucom.​2020.​02.​093. https://​
linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S0925​23122​03028​97
Almatrafi O, Johri A (2022) Improving MOOCs using information from discussion forums: an opinion sum-
marization and suggestion mining approach. IEEE Access 10:15565–15573. https://​doi.​org/​10.​1109/​
ACCESS.​2022.​31492​71. https://​ieeex​plore.​ieee.​org/​docum​ent/​97063​74/
Alyami S, Alhothali A, Jamal A (2022) Systematic literature review of Arabic aspect-based sentiment analy-
sis. J King Saud Univ Comput Inf Sci 34(9):6524–6551. https://​doi.​org/​10.​1016/j.​jksuci.​2022.​07.​001.
https://​linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S1319​15782​20022​82
Amin MM, Mao R, Cambria E et al (2024) A wide evaluation of chatgpt on affective computing tasks. IEEE
Trans Affect Comput 1–9. https://​doi.​org/​10.​1109/​TAFFC.​2024.​34195​93. https://​ieeex​plore.​ieee.​org/​
docum​ent/​10572​294
Asghar MZ, Khan A, Zahra SR et al (2019) Aspect-based opinion mining framework using heuristic pat-
terns. Cluster Comput 22(S3):7181–7199. https://​doi.​org/​10.​1007/​s10586-​017-​1096-9
Baccianella S, Esuli A, Sebastiani F (2010) SentiWordNet 3.0: An enhanced lexical resource for sentiment analysis
and opinion mining. In: Calzolari N, Choukri K, Maegaard B et al (eds) Proceedings of the seventh interna-
tional conference on language resources and evaluation (LREC’10). European Language Resources Asso-
ciation (ELRA), Valletta, Malta. http://​www.​lrec-​conf.​org/​proce​edings/​lrec2​010/​pdf/​769_​Paper.​pdf
Bommasani R, Hudson DA, Adeli E et al (2022) On the opportunities and risks of foundation models.
arXiv:​2108.​07258
Brauwers G, Frasincar F (2023) A survey on aspect-based sentiment classification. ACM Comput Surv
55(4):1–37. https://​doi.​org/​10.​1145/​35030​44
Brown TB, Mann B, Ryder N et al (2020) Language models are few-shot learners. arXiv:​2005.​14165
Cambria E, Poria S, Bajpai R et al (2016) SenticNet 4: a semantic resource for sentiment analysis based on concep-
tual primitives. In: Matsumoto Y, Prasad R (eds) Proceedings of COLING 2016, the 26th international con-
ference on computational linguistics: technical papers. The COLING 2016 Organizing Committee, Osaka,
Japan, pp 2666–2677. https://​aclan​tholo​gy.​org/​C16-​1251
Castellanos M, Dayal U, Hsu M, et al (2011) LCI: A social channel analysis platform for live customer
intelligence. In: Proceedings of the 2011 ACM SIGMOD international conference on management of
data, SIGMOD’11. Association for Computing Machinery, New York, pp 1049–1058. https://​doi.​org/​
10.​1145/​19893​23.​19894​36
Cavalcanti D, Prudêncio R (2017) Aspect-based opinion mining in drug reviews. In: Oliveira E, Gama J,
Vale Z et al (eds) Progress in artificial intelligence, vol 10423. Lecture Notes in Computer Science.
Springer, Cham, pp 815–827. https://​doi.​org/​10.​1007/​978-3-​319-​65340-2_​66
Chauhan GS, Agrawal P, Meena YK (2019) Aspect-based sentiment analysis of students’ feedback to improve
teaching-learning process. In: Satapathy SC, Joshi A (eds) Information and communication technology for
intelligent systems, vol 107. Smart Innovation, Systems and Technologies. Springer Singapore, Singapore,
pp 259–266. https://​doi.​org/​10.​1007/​978-​981-​13-​1747-7_​25
Chebolu SUS, Dernoncourt F, Lipka N et al (2023) Survey of aspect-based sentiment analysis datasets.
arXiv:​2204.​05232
Chen S, Fnu G (2022) Deep learning techniques for aspect based sentiment analysis. In: 2022 14th Inter-
national conference on computer research and development (ICCRD), pp 69–73. https://​doi.​org/​
10.​1109/​ICCRD​54409.​2022.​97304​43. https://​ieeex​plore.​ieee.​org/​docum​ent/​97304​43
Chen Z, Liu B (2014) Topic modeling using topics from many domains, lifelong learning and big data. In:
Proceedings of the 31st International Conference on International Conference on Machine Learning
- Volume 32. JMLR.org, ICML’14, p II–703–II–711. https://​proce​edings.​mlr.​press/​v32/​chenf​14.​html
A systematic review of aspect‑based sentiment analysis: domains,… Page 43 of 51 296

Chen Z, Qian T (2022) Retrieve-and-edit domain adaptation for end2end aspect based sentiment analy-
sis. IEEE/ACM Trans Audio Speech Lang Process 30:659–672. https://​doi.​org/​10.​1109/​TASLP.​
2022.​31460​52. https://​ieeex​plore.​ieee.​org/​docum​ent/​96932​67/
Chen M, Wu W, Zhang Y et al (2021) Combining adversarial training and relational graph attention
network for aspect-based sentiment analysis with BERT. In: 2021 14th international congress
on image and signal processing, BioMedical engineering and informatics (CISP-BMEI), pp 1–6.
https://​doi.​org/​10.​1109/​CISP-​BMEI5​3629.​2021.​96243​84. https://​ieeex​plore.​ieee.​org/​docum​ent/​
96243​84
Chen F, Yang Z, Huang Y (2022) A multi-task learning framework for end-to-end aspect sentiment triplet
extraction. Neurocomputing 479:12–21. https://​doi.​org/​10.​1016/j.​neucom.​2022.​01.​021. https://​linki​
nghub.​elsev​ier.​com/​retri​eve/​pii/​S0925​23122​20004​06
Chu JSG, Evans JA (2021) Slowed canonical progress in large fields of science. Proc Natl Acad Sci
118(41):e2021636118. https://​doi.​org/​10.​1073/​pnas.​20216​36118. https://​pnas.​org/​doi/​full/​10.​1073/​
pnas.​20216​36118
Cortis K, Davis B (2021) Over a decade of social opinion mining: a systematic review. Artif Intell Rev
54(7):4873–4965. https://​doi.​org/​10.​1007/​s10462-​021-​10030-2
Cruz I, Gelbukh AF, Sidorov G (2014) Implicit aspect indicator extraction for aspect based opinion mining.
Int J Comput Linguistics Appl 5(2):135–152. https://​www.​seman​ticsc​holar.​org/​paper/​Impli​cit-​Aspect-​
Indic​ator-​Extra​ction-​for-​Aspect-​Cruz-​Gelbu​kh/​8768f​c3374​b27c0​ac023​f5bf6​0da9a​b5071​4b37e
Dang TV, Hao D, Nguyen N (2024) Vi-AbSQA: multi-task prompt instruction tuning model for Vietnamese
aspect-based sentiment quadruple analysis. ACM Trans Asian Low-Resour Lang Inf Process. https://​
doi.​org/​10.​1145/​36768​86, just Accepted
Da’u A, Salim N, Rabiu I et al (2020) Weighted aspect-based opinion mining using deep learning for recom-
mender system. Expert Syst Appl 140:112871. https://​doi.​org/​10.​1016/j.​eswa.​2019.​112871. https://​
linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S0957​41741​93058​10
Devlin J, Chang M, Lee K et al (2019) BERT: pre-training of deep bidirectional transformers for language under-
standing. In: Burstein J, Doran C, Solorio T (eds) Proceedings of the 2019 conference of the North Ameri-
can chapter of the association for computational linguistics: human language technologies, NAACL-HLT
2019, Minneapolis, MN, USA, June 2–7, 2019, volume 1 (long and short papers). Association for Computa-
tional Linguistics, pp 4171–4186. https://​doi.​org/​10.​18653/​V1/​N19-​1423
Do HH, Prasad P, Maag A et al (2019) Deep learning for aspect-based sentiment analysis: a comparative
review. Expert Syst Appl 118:272–299. https://​doi.​org/​10.​1016/j.​eswa.​2018.​10.​003. https://​www.​
scien​cedir​ect.​com/​scien​ce/​artic​le/​pii/​S0957​41741​83064​56
Dong L, Wei F, Tan C, et al (2014) Adaptive recursive neural network for target-dependent Twitter senti-
ment classification. In: Toutanova K, Wu H (eds) Proceedings of the 52nd Annual Meeting of the
Association for Computational Linguistics (Volume 2: Short Papers). Association for Computational
Linguistics, Baltimore, Maryland, pp 49–54. https://​doi.​org/​10.​3115/​v1/​P14-​2009. https://​aclan​tholo​
gy.​org/​P14-​2009
Dong Q, Li L, Dai D et al (2024) A survey on in-context learning. arXiv:​2301.​00234
Dragoni M, Federici M, Rexha A (2019) An unsupervised aspect extraction strategy for monitoring real-
time reviews stream. Inf Process Manag 56(3):1103–1118. https://​doi.​org/​10.​1016/j.​ipm.​2018.​04.​
010. https://​linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S0306​45731​73051​74
Du C, Wang J, Sun H et al (2021) Syntax-type-aware graph convolutional networks for natural language
understanding. Appl Soft Computi 102:107080. https://​doi.​org/​10.​1016/j.​asoc.​2021.​107080.
https://​linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S1568​49462​10000​3X
Ettaleb M, Barhoumi A, Camelin N et al (2022) Evaluation of weakly-supervised methods for aspect
extraction. Proc Comput Sci 207:2688–2697. https://​doi.​org/​10.​1016/j.​procs.​2022.​09.​327. https://​
linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S1877​05092​20121​69
Fan F, Feng Y, Zhao D (2018) Multi-grained attention network for aspect-level sentiment classification.
In: Conference on empirical methods in natural language processing. https://​api.​seman​ticsc​holar.​
org/​Corpu​sID:​53080​156
Federici M, Dragoni M (2016) A knowledge-based approach for aspect-based opinion mining. In: Sack
H, Dietze S, Tordai A et al (eds) Semantic web challenges, vol 641. Communications in Computer
and Information Science. Springer, Cham, pp 141–152. https://​doi.​org/​10.​1007/​978-3-​319-​46565-
4_​11
Fei H, Chua TS, Li C et al (2023) On the robustness of aspect-based sentiment analysis: rethinking
model, data, and training. ACM Trans Inf Syst 41(2):1–32. https://​doi.​org/​10.​1145/​35642​81.
https://​dl.​acm.​org/​doi/​10.​1145/​35642​81
296 Page 44 of 51 Y. C. Hua et al.

Fei H, Ren Y, Zhang Y et al (2023) Nonautoregressive encoder-decoder neural framework for end-to-end
aspect-based sentiment triplet extraction. IEEE Trans Neural Netw Learn Syst 34(9):5544–5556.
https://​doi.​org/​10.​1109/​TNNLS.​2021.​31294​83. https://​ieeex​plore.​ieee.​org/​docum​ent/​96348​49/
Fernando J, Khodra ML, Septiandri AA (2019) Aspect and opinion terms extraction using double embeddings
and attention mechanism for Indonesian hotel reviews. In: 2019 International conference of advanced infor-
matics: concepts, theory and applications (ICAICTA). IEEE, Yogyakarta, pp 1–6. https://​doi.​org/​10.​1109/​
ICAIC​TA.​2019.​89041​24. https://​ieeex​plore.​ieee.​org/​docum​ent/​89041​24/
FiQA (2018) Financial opinion mining and question answering. https://​sites.​google.​com/​view/​fiqa/​home
Freitas C, Motta E, Milidi´u RL, et al (2014) Sparkling vampire...lol! annotating opinions in a book review
corpus. New language technologies and linguistic research: a two-way Road pp 128–146. https://​
www.​resea​rchga​te.​net/​publi​cation/​27183​6545_​Spark​ling_​Vampi​re_​lol_​Annot​ating_​Opini​ons_​in_a_​
Book_​Review_​Corpus
Fukumoto F, Sugiyama H, Suzuki Y et al (2016) Exploiting guest preferences with aspect-based sentiment analy-
sis for hotel recommendation. In: Fred A, Dietz JL, Aveiro D et al (eds) Knowledge discovery, knowledge
engineering and knowledge management, vol 631. Communications in Computer and Information Science.
Springer, Cham, pp 31–46. https://​doi.​org/​10.​1007/​978-3-​319-​52758-1_3
Ganganwar V, Rajalakshmi R (2019) Implicit aspect extraction for sentiment analysis: a survey of recent
approaches. Proc Comput Sci 165:485–491. https://​doi.​org/​10.​1016/j.​procs.​2020.​01.​010. https://​www.​scien​
cedir​ect.​com/​scien​ce/​artic​le/​pii/​S1877​05092​03001​81, 2nd International Conference on Recent Trends in
Advanced Computing ICRTAC-DISRUP - TIV INNOVATION, 2019 November 11–12, 2019
Ganu G, Elhadad N, Marian A (2009) Beyond the stars: improving rating predictions using review text
content. In: International workshop on the web and databases. https://​api.​seman​ticsc​holar.​org/​
Corpu​sID:​18345​070
García-Pablos A, Cuadros M, Rigau G (2018) W2vlda: Almost unsupervised system for aspect based
sentiment analysis. Expert Syst Appl 91:127–137. https://​doi.​org/​10.​1016/j.​eswa.​2017.​08.​049.
https://​linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S0957​41741​73059​61
Go A (2009) Twitter sentiment classification using distant supervision. CS224N Project Report, Stanford,
1-12. https://​api.​seman​ticsc​holar.​org/​Corpu​sID:​18635​269
Gojali S, Khodra ML (2016) Aspect based sentiment analysis for review rating prediction. In: 2016
International conference on advanced informatics: concepts, theory and application (ICAICTA).
IEEE, Penang, pp 1–6. https://​doi.​org/​10.​1109/​ICAIC​TA.​2016.​78031​10. http://​ieeex​plore.​ieee.​
org/​docum​ent/​78031​10/
Gong Z, Li B (2022) Emotional text generation with hard constraints. In: 2022 4th International confer-
ence on frontiers technology of information and computer (ICFTIC), pp 68–73. https://​doi.​org/​10.​
1109/​ICFTI​C57696.​2022.​10075​091. https://​ieeex​plore.​ieee.​org/​docum​ent/​10075​091
Gui L, He Y (2021) Understanding patient reviews with minimum supervision. Artif Intell Med
120:102160. https://​doi.​org/​10.​1016/j.​artmed.​2021.​102160. https://​www.​scien​cedir​ect.​com/​scien​
ce/​artic​le/​pii/​S0933​36572​10015​36
Gunes O (2016) Aspect term and opinion target extraction from web product reviews using semi-Markov
conditional random fields with word embeddings as features. In: Proceedings of the 6th interna-
tional conference on web intelligence, mining and semantics. ACM, Nìmes, pp 1–5. https://​doi.​
org/​10.​1145/​29128​45.​29368​09
Guo L, Jiang S, Du W et al (2018) Recurrent neural CRF for aspect term extraction with dependency
transmission. In: Zhang M, Ng V, Zhao D et al (eds) Natural language processing and Chinese
computing, vol 11108. Lecture Notes in Computer Science. Springer, Cham, pp 378–390. https://​
doi.​org/​10.​1007/​978-3-​319-​99495-6_​32
He R, McAuley J (2016) Ups and downs: Modeling the visual evolution of fashion trends with one-class
collaborative filtering. In: Proceedings of the 25th International Conference on World Wide Web.
International World Wide Web Conferences Steering Committee, Republic and Canton of Geneva,
CHE, WWW ’16, p 507–517. https://​doi.​org/​10.​1145/​28724​27.​28830​37
He R, Lee WS, Ng HT et al (2019) An interactive multi-task learning network for end-to-end aspect-
based sentiment analysis. In: Proceedings of the 57th annual meeting of the association for com-
putational linguistics. Association for Computational Linguistics, Florence, pp 504–515. https://​
doi.​org/​10.​18653/​v1/​P19-​1048. https://​www.​aclweb.​org/​antho​logy/​P19-​1048
Hoang CD, Dinh QV, Tran NH (2022) Aspect-category-opinion-sentiment extraction using generative
transformer model. In: 2022 RIVF international conference on computing and communication
technologies (RIVF), pp 1–6. https://​doi.​org/​10.​1109/​RIVF5​5975.​2022.​10013​820. https://​ieeex​
plore.​ieee.​org/​docum​ent/​10013​820
Hoti MH, Ajdari J, Hamiti M et al (2022) Text mining, clustering and sentiment analysis: a system-
atic literature review. In: 2022 11th Mediterranean conference on embedded computing (MECO).
A systematic review of aspect‑based sentiment analysis: domains,… Page 45 of 51 296

IEEE, Budva, Montenegro, pp 1–6. https://​doi.​org/​10.​1109/​MECO5​5406.​2022.​97972​03. https://​

ieeex​plore.​ieee.​org/​docum​ent/​97972​03/
Howard P, Ma A, Lal V et al (2022) Cross-domain aspect extraction using transformers augmented with
knowledge graphs. In: Proceedings of the 31st ACM international conference on information &
knowledge management. ACM, Atlanta, pp 780–790. https://​doi.​org/​10.​1145/​35118​08.​35572​75
Huan H, He Z, Xie Y et al (2022) A multi-task dual-encoder framework for aspect sentiment triplet extrac-
tion. IEEE Access 10:103187–103199. https://​doi.​org/​10.​1109/​ACCESS.​2022.​32101​80. https://​ieeex​
plore.​ieee.​org/​docum​ent/​99036​19/
Hu M, Liu B (2004a) Mining and summarizing customer reviews. In: Proceedings of the tenth ACM SIG-
KDD international conference on Knowledge discovery and data mining. ACM, Seattle, pp 168–177.
https://​doi.​org/​10.​1145/​10140​52.​10140​73
Hu M, Liu B (2004b) Mining opinion features in customer reviews. In: AAAI conference on artificial intel-
ligence. https://​api.​seman​ticsc​holar.​org/​Corpu​sID:​57248​60
J AK, Trueman TE, Cambria E (2021) A convolutional stacked bidirectional LSTM with a multiplicative
attention mechanism for aspect category and sentiment detection. Cogn Comput 13(6):1423–1432.
https://​doi.​org/​10.​1007/​s12559-​021-​09948-0
Jiang Q, Chen L, Xu R et al (2019) A challenge dataset and effective models for aspect-based sentiment
analysis. In: Proceedings of the 2019 conference on empirical methods in natural language processing
and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). Asso-
ciation for Computational Linguistics, Hong Kong, pp 6279–6284. https://​doi.​org/​10.​18653/​v1/​D19-​
1654. https://​www.​aclweb.​org/​antho​logy/​D19-​1654
Jo Y, Oh AH (2011) Aspect and sentiment unification model for online review analysis. In: Proceedings of
the Fourth ACM International Conference on Web Search and Data Mining. Association for Com-
puting Machinery, New York, NY, USA, WSDM ’11, p 815–824. https://​doi.​org/​10.​1145/​19358​26.​
19359​32
Kang T, Kim S, Yun H et al (2022) Gated relational encoder-decoder model for target-oriented opinion
word extraction. IEEE Access 10:130507–130517. https://​doi.​org/​10.​1109/​ACCESS.​2022.​32288​35.
https://​ieeex​plore.​ieee.​org/​docum​ent/​99826​01
Kitchenham BA, Charters S (2007) Guidelines for performing systematic literature reviews in software
engineering. https://​www.​elsev​ier.​com/__​data/​promis_​misc/​52544​4syst​emati​crevi​ewsgu​ide.​pdf.
Backup Publisher: Keele University and Durham University Joint Report
Krishnakumari K, Sivasankar E (2018) Scalable aspect-based summarization in the hadoop environment.
In: Aggarwal VB, Bhatnagar V, Mishra DK (eds) Big data analytics, vol 654. Advances in Intelligent
Systems and Computing. Springer Singapore, Singapore, pp 439–449. https://​doi.​org/​10.​1007/​978-​
981-​10-​6620-7_​42
Kumar A, Gupta D (2021) Sentiment analysis as a restricted NLP problem:. In: Pinarbasi F, Taskiran MN
(eds) Advances in business information systems and analytics. IGI Global, pp 65–96. https://​doi.​org/​
10.​4018/​978-1-​7998-​4240-8.​ch004. http://​servi​ces.​igi-​global.​com/​resol​vedoi/​resol​ve.​aspx?​doi=​10.​
4018/​978-1-​7998-​4240-8.​ch004
Kumar A, Seth S, Gupta S et al (2022) Sentic computing for aspect-based opinion summarization using
multi-head attention with feature pooled pointer generator network. Cogn Comput 14(1):130–148.
https://​doi.​org/​10.​1007/​s12559-​021-​09835-8
Lee SK, Kim JH (2023) Sener: Sentiment element named entity recognition for aspect-based sentiment
analysis. In: ICASSP 2023—2023 IEEE international conference on acoustics, speech and signal pro-
cessing (ICASSP), pp 1–5. https://​doi.​org/​10.​1109/​ICASS​P49357.​2023.​10095​101. https://​ieeex​plore.​
ieee.​org/​docum​ent/​10095​101
Lee C, Lee H, Kim K et al (2024) An efficient fine-tuning of generative language model for aspect-based
sentiment analysis. In: 2024 IEEE international conference on consumer electronics (ICCE), pp 1–4.
https://​doi.​org/​10.​1109/​ICCE5​9016.​2024.​10444​216. https://​ieeex​plore.​ieee.​org/​docum​ent/​10444​216
Lewis M, Liu Y, Goyal N et al (2020) BART: Denoising sequence-to-sequence pre-training for natural lan-
guage generation, translation, and comprehension. In: Jurafsky D, Chai J, Schluter N et al (eds) Pro-
ceedings of the 58th annual meeting of the association for computational linguistics. Association for
Computational Linguistics, Online, pp 7871–7880. https://​doi.​org/​10.​18653/​v1/​2020.​acl-​main.​703.
https://​aclan​tholo​gy.​org/​2020.​acl-​main.​703
Li X, Wang B, Li L et al (2020) Deep2s: Improving aspect extraction in opinion mining with deep seman-
tic representation. IEEE Access 8:104026–104038. https://​doi.​org/​10.​1109/​ACCESS.​2020.​29996​73.
https://​ieeex​plore.​ieee.​org/​docum​ent/​91071​47/
Li Z, Li L, Zhou A et al (2021) JTSG: A joint term-sentiment generator for aspect-based sentiment analysis.
Neurocomputing 459:1–9. https://​doi.​org/​10.​1016/j.​neucom.​2021.​06.​045. https://​www.​scien​cedir​ect.​
com/​scien​ce/​artic​le/​pii/​S0925​23122​10096​93
296 Page 46 of 51 Y. C. Hua et al.

Li J, Zhao Y, Jin Z et al (2022a) SK2: Integrating implicit sentiment knowledge and explicit syntax knowl-
edge for aspect-based sentiment analysis. In: Proceedings of the 31st ACM international conference
on information & knowledge management. ACM, Atlanta, pp 1114–1123. https://​doi.​org/​10.​1145/​
35118​08.​35574​52
Li Y, Lin Y, Lin Y et al (2022) A span-sharing joint extraction framework for harvesting aspect sentiment
triplets. Knowl Based Syst 242:108366. https://​doi.​org/​10.​1016/j.​knosys.​2022.​108366. https://​linki​
nghub.​elsev​ier.​com/​retri​eve/​pii/​S0950​70512​20013​81
Li Y, Wang C, Lin Y et al (2022) Span-based relational graph transformer network for aspect-opinion pair
extraction. Knowl Inf Syst 64(5):1305–1322. https://​doi.​org/​10.​1007/​s10115-​022-​01675-8
Li S, Zhang Y, Lan Y et al (2023) From implicit to explicit: a simple generative method for aspect-category-
opinion-sentiment quadruple extraction. In: 2023 international joint conference on neural networks
(IJCNN), pp 1–8. https://​doi.​org/​10.​1109/​IJCNN​54540.​2023.​10191​098. https://​ieeex​plore.​ieee.​org/​
docum​ent/​10191​098
Liang B, Su H, Gui L et al (2022) Aspect-based sentiment analysis via affective knowledge enhanced graph
convolutional networks. Knowl Based Syst 235:107643. https://​doi.​org/​10.​1016/j.​knosys.​2021.​
107643. https://​linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S0950​70512​10090​59
Ligthart A, Catal C, Tekinerdogan B (2021) Systematic reviews in sentiment analysis: a tertiary study. Artif
Intell Rev 54(7):4997–5053. https://​doi.​org/​10.​1007/​s10462-​021-​09973-3
Lil Z, Yang Z, Li X et al (2023) Two-stage aspect sentiment quadruple prediction based on MRC and text
generation. In: 2023 IEEE International conference on systems, man, and cybernetics (SMC), pp
2118–2125. https://​doi.​org/​10.​1109/​SMC53​992.​2023.​10394​369. https://​ieeex​plore-​ieee-​org.​ezpro​xy.​
auckl​and.​ac.​nz/​docum​ent/​10394​369
Lim KW, Buntine W (2014) Twitter opinion topic model: extracting product opinions from tweets by lever-
aging hashtags and sentiment lexicon. In: Proceedings of the 23rd ACM international conference on
conference on information and knowledge management. ACM, Shanghai, pp 1319–1328. https://​doi.​
org/​10.​1145/​26618​29.​26620​05
Lin B, Cassee N, Serebrenik A et al (2022) Opinion mining for software development: a systematic litera-
ture review. ACM Trans Softw Eng Methodol 31(3):1–41. https://​doi.​org/​10.​1145/​34903​88
Liu Q, Gao Z, Liu B, et al (2015) Automated rule selection for aspect extraction in opinion mining. In:
Proceedings of the 24th International Conference on Artificial Intelligence. AAAI Press, IJCAI’ 15, p
1291–1297. https://​doi.​org/​10.​5555/​28324​15.​28324​29
Liu Y, Ott M, Goyal N et al (2019) Roberta: a robustly optimized BERT pretraining approach. arXiv:​1907.​
11692
Liu H, Chatterjee I, Zhou M et al (2020) Aspect-based sentiment analysis: a survey of deep learning meth-
ods. IEEE Trans Comput Soc Syst 7(6):1358–1375. https://​doi.​org/​10.​1109/​TCSS.​2020.​30333​02.
https://​ieeex​plore.​ieee.​org/​docum​ent/​92601​62/
Liu J, Chen T, Guo H et al (2024) Exploiting duality in aspect sentiment triplet extraction with sequential
prompting. IEEE Trans Knowl Data Eng 1–12. https://​doi.​org/​10.​1109/​TKDE.​2024.​33913​81. https://​
ieeex​plore.​ieee.​org/​docum​ent/​10505​831
López D, Arco L (2019) Multi-domain aspect extraction based on deep and lifelong learning. In: Nyström
I, Hernández Heredia Y, Milián Núñez V (eds) Progress in pattern recognition, image analysis, com-
puter vision, and applications, vol 11896. Lecture Notes in Computer Science. Springer, Cham, pp
556–565. https://​doi.​org/​10.​1007/​978-3-​030-​33904-3_​52
Luo H, Li T, Liu B et al (2019) Improving aspect term extraction with bidirectional dependency tree repre-
sentation. IEEE/ACM Trans Audio Speech Lang Process 27(7):1201–1212. https://​doi.​org/​10.​1109/​
TASLP.​2019.​29130​94. https://​ieeex​plore.​ieee.​org/​docum​ent/​86983​40/
Ma Y, Chen G, Wei Q (2017) Finding users preferences from large-scale online reviews for personalized
recommendation. Electron Commer Res 17(1):3–29. https://​doi.​org/​10.​1007/​s10660-​016-​9240-9
Maitama JZ, Idris N, Abdi A et al (2020) A systematic review on implicit and explicit aspect extraction in
sentiment analysis. IEEE Access 8:194166–194191. https://​doi.​org/​10.​1109/​ACCESS.​2020.​30312​17.
https://​ieeex​plore.​ieee.​org/​docum​ent/​92344​64/
Manning CD (2022) Human language understanding & reasoning. Daedalus 151(2):127–138. https://​doi.​
org/​10.​1162/​daed_a_​01905. https://​direct.​mit.​edu/​daed/​artic​le/​151/2/​127/​110621/​Human-​Langu​age-​
Under​stand​ing-​amp-​Reaso​ning
Marstawi A, Sharef NM, Aris TNM, et al (2017) Ontology-based aspect extraction for an improved senti-
ment analysis in summarization of product reviews. In: Proceedings of the 8th international confer-
ence on computer modeling and simulation, ICCMS’17. Association for Computing Machinery, New
York, pp 100–104. https://​doi.​org/​10.​1145/​30363​31.​30363​62
A systematic review of aspect‑based sentiment analysis: domains,… Page 47 of 51 296

McAuley J, Leskovec J, Jurafsky D (2012) Learning attitudes and attributes from multi-aspect reviews. In:
Proceedings of the 2012 IEEE 12th International Conference on Data Mining. IEEE Computer Soci-
ety, USA, ICDM ’12, p 1020–1025. https://​doi.​org/​10.​1109/​ICDM.​2012.​110
McAuley J, Targett C, Shi Q, Van Den Hengel A (2015) Image-based recommendations on styles and sub-
stitutes. In Proceedings of the 38th international ACM SIGIR conference on research and develop-
ment in information retrieval Association for Computing Machinery, New York, NY, USA, SIGIR’15,
p 43–52. https://​doi.​org/​10.​1145/​27664​62.​27677​55
Mitchell M, Aguilar J, Wilson T, et al (2013) Open domain targeted sentiment. In: Yarowsky D, Baldwin T,
Korhonen A, et al (eds) Proceedings of the 2013 Conference on Empirical Methods in Natural Lan-
guage Processing. Association for Computational Linguistics, Seattle, Washington, USA, pp 1643–
1654. https://​aclan​tholo​gy.​org/​D13-​1171
Mughal N, Mujtaba G, Shaikh S et al (2024) Comparative analysis of deep natural networks and large lan-
guage models for aspect-based sentiment analysis. IEEE Access 12:60943–60959. https://​doi.​org/​10.​
1109/​ACCESS.​2024.​33869​69. https://​ieeex​plore.​ieee.​org/​docum​ent/​10504​711
Nawaz A, Awan AA, Ali T et al (2020) Product’s behaviour recommendations using free text: an aspect
based sentiment analysis approach. Clust Comput 23(2):1267–1279. https://​doi.​org/​10.​1007/​
s10586-​019-​02995-1
Nazir A, Rao Y (2022) IAOTP: An interactive end-to-end solution for aspect-opinion term pairs extraction.
In: Proceedings of the 45th international ACM SIGIR conference on research and development in
information retrieval. ACM, Madrid, pp 1588–1598. https://​doi.​org/​10.​1145/​34774​95.​35320​85
Nazir A, Rao Y, Wu L et al (2022) IAF-LG: an interactive attention fusion network with local and global
perspective for aspect-based sentiment analysis. IEEE Trans Affect Comput 13(4):1730–1742.
https://​doi.​org/​10.​1109/​TAFFC.​2022.​32082​16. https://​ieeex​plore.​ieee.​org/​docum​ent/​98969​31/
Nazir A, Rao Y, Wu L et al (2022) Issues and challenges of aspect-based sentiment analysis: a compre-
hensive survey. IEEE Trans Affect Comput 13(2):845–863. https://​doi.​org/​10.​1109/​TAFFC.​2020.​
29703​99. https://​ieeex​plore.​ieee.​org/​docum​ent/​89762​52/
Obiedat R, Al-Darras D, Alzaghoul E et al (2021) Arabic aspect-based sentiment analysis: a systematic
literature review. IEEE Access 9:152628–152645. https://​doi.​org/​10.​1109/​ACCESS.​2021.​31271​
40. https://​ieeex​plore.​ieee.​org/​docum​ent/​96112​71/
OpenAI (2023) Chatgpt (mar 14 version) [large language model]. https://​chat.​openai.​com/​chat
Pathan AF, Prakash C (2022) Cross-domain aspect detection and categorization using machine learning
for aspect-based opinion mining. Int J Inf Manag Data Insights 2(2):100099. https://​doi.​org/​10.​
1016/j.​jjimei.​2022.​100099. https://​www.​scien​cedir​ect.​com/​scien​ce/​artic​le/​pii/​S2667​09682​20004​
28
Peng H, Ma Y, Li Y, et al (2018) Learning multi-grained aspect target sequence for chinese sentiment analy-
sis. Knowledge-Based Syst 148:167–176. https://​doi.​org/​10.​1016/j.​knosys.​2018.​02.​034. https://​www.​
scien​cedir​ect.​com/​scien​ce/​artic​le/​pii/​S0950​70511​83009​72
Phan MH, Ogunbona PO (2020) Modelling context and syntactical features for aspect-based sentiment
analysis. In: Proceedings of the 58th annual meeting of the association for computational linguis-
tics. Association for Computational Linguistics, Online, pp 3211–3220. https://​doi.​org/​10.​18653/​
v1/​2020.​acl-​main.​293. https://​www.​aclweb.​org/​antho​logy/​2020.​acl-​main.​293
Pontiki M, Galanis D, Pavlopoulos J et al (2014) SemEval-2014 task 4: aspect based sentiment analysis.
In: Nakov P, Zesch T (eds) Proceedings of the 8th international workshop on semantic evaluation
(SemEval 2014). Association for Computational Linguistics, Dublin, pp 27–35. https://​doi.​org/​10.​
3115/​v1/​S14-​2004. https://​aclan​tholo​gy.​org/​S14-​2004
Pontiki M, Galanis D, Papageorgiou H et al (2015) SemEval-2015 task 12: aspect based sentiment analy-
sis. In: Nakov P, Zesch T, Cer D et al (eds) Proceedings of the 9th international workshop on
semantic evaluation (SemEval 2015). Association for Computational Linguistics, Denver, pp 486–
495. https://​doi.​org/​10.​18653/​v1/​S15-​2082. https://​aclan​tholo​gy.​org/​S15-​2082
Pontiki M, Galanis D, Papageorgiou H et al (2016) SemEval-2016 task 5: aspect based sentiment analy-
sis. In: Bethard S, Carpuat M, Cer D et al (eds) Proceedings of the 10th international workshop on
semantic evaluation (SemEval-2016). Association for Computational Linguistics, San Diego, pp
19–30. https://​doi.​org/​10.​18653/​v1/​S16-​1002. https://​aclan​tholo​gy.​org/​S16-​1002
Poria S, Chaturvedi I, Cambria E et al (2016) Sentic LDA: Improving on LDA with semantic similarity
for aspect-based sentiment analysis. In: 2016 International joint conference on neural networks
(IJCNN). IEEE, Vancouver, pp 4465–4473. https://​doi.​org/​10.​1109/​IJCNN.​2016.​77277​84. http://​
ieeex​plore.​ieee.​org/​docum​ent/​77277​84/
Prather J, Becker BA, Craig M et al (2020) What do we think we think we are doing?: Metacognition
and self-regulation in programming. In: Proceedings of the 2020 ACM conference on international
296 Page 48 of 51 Y. C. Hua et al.

computing education research. ACM, Virtual Event New Zealand, pp 2–13. https://​doi.​org/​10.​
1145/​33727​82.​34062​63
Presannakumar K, Mohamed A (2021) An enhanced method for review mining using n-gram
approaches. In: Raj JS, Iliyasu AM, Bestak R et al (eds) Innovative data communication technolo-
gies and application, vol 59. Lecture Notes on Data Engineering and Communications Technolo-
gies. Springer Singapore, Singapore, pp 615–626. https://​doi.​org/​10.​1007/​978-​981-​15-​9651-3_​51
Radford A, Wu J, Child R et al (2019) Language models are unsupervised multitask learners. https://​api.​
seman​ticsc​holar.​org/​Corpu​sID:​16002​5533
Raffel C, Shazeer N, Roberts A et al (2020) Exploring the limits of transfer learning with a unified text-
to-text transformer. J Mach Learn Res 21(1). https://​dl.​acm.​org/​doi/​abs/​10.​5555/​34557​16.​34558​56
Rahman MA, Kumar Dey E (2018) Datasets for aspect-based sentiment analysis in bangla and its baseline
evaluation. Data 3(2). https://​doi.​org/​10.​3390/​data3​020015. https://​www.​mdpi.​com/​2306-​5729/3/​2/​15
Rana TA, Cheah YN (2016) Aspect extraction in sentiment analysis: comparative analysis and survey.
Artif Intell Rev 46:459–483. https://​api.​seman​ticsc​holar.​org/​Corpu​sID:​24401​592
Rani S, Kumar P (2019) A journey of Indian languages over sentiment analysis: a systematic review.
Artif Intell Rev 52(2):1415–1462. https://​doi.​org/​10.​1007/​s10462-​018-​9670-y
Ruskanda FZ, Widyantoro DH, Purwarianti A (2019) Sequential covering rule learning for language
rule-based aspect extraction. In: 2019 International conference on advanced computer science and
information systems (ICACSIS). IEEE, Bali, pp 229–234. https://​doi.​org/​10.​1109/​ICACS​IS477​36.​
2019.​89797​43. https://​ieeex​plore.​ieee.​org/​docum​ent/​89797​43/
Sabeeh A, Dewang RK (2019) Comparison, classification and survey of aspect based sentiment analysis.
In: Luhach AK, Singh D, Hsiung PA et al (eds) Advanced informatics for computing research.
Springer Singapore, Singapore, pp 612–629. https://​doi.​org/​10.​1007/​978-​981-​13-​3140-4_​55
Saeidi M, Bouchard G, Liakata M, et al (2016) SentiHood: Targeted aspect based sentiment analysis dataset
for urban neighbourhoods. In: Matsumoto Y, Prasad R (eds) Proceedings of COLING 2016, the 26th
International Conference on Computational Linguistics: Technical Papers. The COLING 2016 Organ-
izing Committee, Osaka, Japan, pp 1546–1556. https://​aclan​tholo​gy.​org/​C16-​1146
Sanders NJ (2011) Sanders-twitter sentiment corpus. Sanders Analytics LLC
Satyarthi S, Sharma S (2023) Identification of effective deep learning approaches for classifying senti-
ments at aspect level in different domain. In: 2023 IEEE International conference on paradigm
shift in information technologies with innovative applications in global scenario (ICPSITIAGS),
pp 496–508. https://​doi.​org/​10.​1109/​ICPSI​TIAGS​59213.​2023.​10527​695. https://​ieeex​plore-​ieee-​
org.​ezpro​xy.​auckl​and.​ac.​nz/​docum​ent/​10527​695
Sharma A, Shekhar H (2020) Intelligent learning based opinion mining model for governmental decision
making. Proc Comput Sci 173:216–224. https://​doi.​org/​10.​1016/j.​procs.​2020.​06.​026. https://​linki​
nghub.​elsev​ier.​com/​retri​eve/​pii/​S1877​05092​03153​01
Socher R, Perelygin A, Wu J, et al (2013) Recursive deep models for semantic compositionality over a
sentiment treebank. In: Yarowsky D, Baldwin T, Korhonen A, et al (eds) Proceedings of the 2013
Conference on Empirical Methods in Natural Language Processing. Association for Computational
Linguistics, Seattle, Washington, USA, pp 1631–1642. https://​aclan​tholo​gy.​org/​D13-​1170
Soni PK, Rambola R (2022) A survey on implicit aspect detection for sentiment analysis: terminology,
issues, and scope. IEEE Access 10:63932–63957. https://​doi.​org/​10.​1109/​ACCESS.​2022.​31832​05.
https://​ieeex​plore.​ieee.​org/​docum​ent/​97965​23
Suchrady RZ, Purwarianti A (2023) Indo LEGO-ABSA: a multitask generative aspect based sentiment anal-
ysis for Indonesian language. In: 2023 International conference on electrical engineering and infor-
matics (ICEEI), pp 1–6. https://​doi.​org/​10.​1109/​ICEEI​59426.​2023.​10346​852. https://​ieeex​plore-​ieee-​
org.​ezpro​xy.​auckl​and.​ac.​nz/​docum​ent/​10346​852
Sutskever I, Vinyals O, Le QV (2014) Sequence to sequence learning with neural networks. In: Proceedings
of the 27th international conference on neural information processing systems, NIPS’14, vol 2. MIT
Press, Montreal, pp 3104–3112
Su H, Wang X, Li J et al (2024) Enhanced implicit sentiment understanding with prototype learning and
demonstration for aspect-based sentiment analysis. IEEE Trans Comput Soc Syst 1–16. https://​doi.​
org/​10.​1109/​TCSS.​2024.​33681​71. https://​ieeex​plore-​ieee-​org.​ezpro​xy.​auckl​and.​ac.​nz/​docum​ent/​
10584​152
Team TPD (2023) pandas-dev/pandas: Pandas. https://​doi.​org/​10.​5281/​ZENODO.​35091​34. https://​zenodo.​
org/​record/​35091​34
Toprak C, Jakob N, Gurevych I (2010) Sentence and expression level annotation of opinions in user-gen-
erated discourse. In: Proceedings of the 48th Annual Meeting of the Association for Computational
Linguistics. Association for Computational Linguistics, USA, ACL ’10, p 575–584. https://​doi.​org/​
10.​5555/​18586​81.​18587​40
A systematic review of aspect‑based sentiment analysis: domains,… Page 49 of 51 296

Tran TU, Hoang HTT, Huynh HX (2020) Bidirectional independently long short-term memory and con-
ditional random field integrated model for aspect extraction in sentiment analysis. In: Satapathy SC,
Bhateja V, Nguyen BL et al (eds) Frontiers in intelligent computing: theory and applications, vol
1014. Advances in Intelligent Systems and Computing. Springer Singapore, Singapore, pp 131–140.
https://​doi.​org/​10.​1007/​978-​981-​13-​9920-6_​14
Tubishat M, Idris N, Abushariah M (2021) Explicit aspects extraction in sentiment analysis using optimal
rules combination. Futur Gener Comput Syst 114:448–480. https://​doi.​org/​10.​1016/j.​future.​2020.​08.​
019. https://​linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S0167​739X1​93308​1X
Vasanthi A, Kumar H, Karanraj R (2022) An RL approach for ABSA using transformers. In: 2022 6th Inter-
national conference on trends in electronics and informatics (ICOEI), pp 354–361. https://​doi.​org/​10.​
1109/​ICOEI​53556.​2022.​97769​15. https://​ieeex​plore.​ieee.​org/​docum​ent/​97769​15
Vaswani A, Shazeer N, Parmar N, et al (2017) Attention is all you need. In: Proceedings of the 31st inter-
national conference on neural information processing systems, NIPS’17. Curran Associates Inc., Red
Hook, pp 6000–6010. https://​doi.​org/​10.​5555/​32952​22.​32953​49
Wang H, Lu Y, Zhai C (2010) Latent aspect rating analysis on review text data: a rating regression approach.
In: Proceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining. Association for Computing Machinery, New York, NY, USA, KDD ’10, p 783–792.
https://​doi.​org/​10.​1145/​18358​04.​18359​03
Wang H, Lu Y, Zhai C (2011) Latent aspect rating analysis without aspect keyword supervision. In: Pro-
ceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining. Association for Computing Machinery, New York, NY, USA, KDD ’11, p 618–626. https://​
doi.​org/​10.​1145/​20204​08.​20205​05
Wang Y, Huang Y, Wang M (2017) Aspect-based rating prediction on reviews using sentiment strength
analysis. In: Benferhat S, Tabia K, Ali M (eds) Advances in artificial intelligence: from theory to
practice, vol 10351. Lecture Notes in Computer Science. Springer, Cham, pp 439–447. https://​doi.​
org/​10.​1007/​978-3-​319-​60045-1_​45
Wang W, Pan SJ, Dahlmeier D (2018) Memory networks for fine-grained opinion mining. Artif Intell
265:1–17. https://​doi.​org/​10.​1016/j.​artint.​2018.​09.​002. https://​linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​
S0004​37021​83059​9X
Wang J, Xu B, Zu Y (2021a) Deep learning for aspect-based sentiment analysis. In: 2021 International con-
ference on machine learning and intelligent systems engineering (MLISE), pp 267–271. https://​doi.​
org/​10.​1109/​MLISE​54096.​2021.​00056. https://​ieeex​plore.​ieee.​org/​docum​ent/​96117​05
Wang L, Zong B, Liu Y et al (2021b) Aspect-based sentiment classification via reinforcement learning. In:
2021 IEEE international conference on data mining (ICDM), pp 1391–1396. https://​doi.​org/​10.​1109/​
ICDM5​1629.​2021.​00177. https://​ieeex​plore.​ieee.​org/​docum​ent/​96791​12
Wang X, Liu P, Zhu Z et al (2022) Interactive double graph convolutional networks for aspect-based sen-
timent analysis. In: 2022 International joint conference on neural networks (IJCNN). IEEE, Padua,
Italy, pp 1–7. https://​doi.​org/​10.​1109/​IJCNN​55064.​2022.​98929​34. https://​ieeex​plore.​ieee.​org/​docum​
ent/​98929​34/
Wang Z, Xia R, Yu J (2024) Unified ABSA via annotation-decoupled multi-task instruction tuning. IEEE
Trans Knowl Data Eng 1–13. https://​doi.​org/​10.​1109/​TKDE.​2024.​33928​36. https://​ieeex​plore.​ieee.​
org/​docum​ent/​10507​027
Wankhade M, Rao ACS, Kulkarni C (2022) A survey on sentiment analysis methods, applications, and chal-
lenges. Artif Intel Rev 55(7):5731–5780. https://​doi.​org/​10.​1007/​s10462-​022-​10144-1
Wikipedia (2023) SemEval. https://​en.​wikip​edia.​org/​wiki/​SemEv​al
William, Khodra ML (2022) Generative opinion triplet extraction using pretrained language model. In:
2022 9th International conference on advanced informatics: concepts, theory and applications
(ICAICTA), pp 1–6. https://​doi.​org/​10.​1109/​ICAIC​TA564​49.​2022.​99330​04. https://​ieeex​plore.​
ieee.​org/​docum​ent/​99330​04
Wu S, Fei H, Ren Y et al (2021) High-order pair-wise aspect and opinion terms extraction with edge-
enhanced syntactic graph convolution. IEEE/ACM Trans Audio Speech Lang Process 29:2396–
2406. https://​doi.​org/​10.​1109/​TASLP.​2021.​30956​72. https://​ieeex​plore.​ieee.​org/​docum​ent/​94781​
83/
Xing X, Jin Z, Jin D et al (2020) Tasty burgers, soggy fries: probing aspect robustness in aspect-based senti-
ment analysis. In: Proceedings of the 2020 conference on empirical methods in natural language pro-
cessing (EMNLP). Association for Computational Linguistics, Online, pp 3594–3605. https://​doi.​org/​
10.​18653/​v1/​2020.​emnlp-​main.​292. https://​www.​aclweb.​org/​antho​logy/​2020.​emnlp-​main.​292
Xu K, Zhao H, Liu T (2020) Aspect-specific heterogeneous graph convolutional network for aspect-
based sentiment classification. IEEE Access 8:139346–139355. https://​doi.​org/​10.​1109/​ACCESS.​
2020.​30126​37. https://​ieeex​plore.​ieee.​org/​docum​ent/​91520​16/
296 Page 50 of 51 Y. C. Hua et al.

Xu Q, Zhu L, Dai T et al (2020) Non-negative matrix factorization for implicit aspect identification. J
Ambient Intell Humaniz Comput 11(7):2683–2699. https://​doi.​org/​10.​1007/​s12652-​019-​01328-9
Yan K, Tang L, Wu M et al (2023) Aspect-based sentiment analysis method using text generation.
In: Proceedings of the 2023 7th international conference on big data and internet of things,
BDIOT’23. Association for Computing Machinery, New York, pp 156–161. https://​doi.​org/​10.​
1145/​36176​95.​36177​09
Yauris K, Khodra ML (2017) Aspect-based summarization for game review using double propagation.
In: 2017 International conference on advanced informatics, concepts, theory, and applications
(ICAICTA). IEEE, Denpasar, pp 1–6. https://​doi.​org/​10.​1109/​ICAIC​TA.​2017.​80909​97. http://​
ieeex​plore.​ieee.​org/​docum​ent/​80909​97/
You L, Han F, Peng J et al (2022) ASK-RoBERTa: a pretraining model for aspect-based sentiment clas-
sification via sentiment knowledge mining. Knowl Based Syst 253:109511. https://​doi.​org/​10.​
1016/j.​knosys.​2022.​109511. https://​linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S0950​70512​20075​84
Yu C, Wu T, Li J et al (2023a) Syngen: A syntactic plug-and-play module for generative aspect-based
sentiment analysis. In: ICASSP 2023–2023 IEEE international conference on acoustics, speech
and signal processing (ICASSP), pp 1–5. https://​doi.​org/​10.​1109/​ICASS​P49357.​2023.​10094​591.
https://​ieeex​plore.​ieee.​org/​docum​ent/​10094​591
Yu Y, Zhao M, Zhou S (2023b) Boosting aspect sentiment quad prediction by data augmentation and
self-training. In: 2023 International joint conference on neural networks (IJCNN), pp 1–8. https://​
doi.​org/​10.​1109/​IJCNN​54540.​2023.​10191​634. https://​ieeex​plore.​ieee.​org/​docum​ent/​10191​634
Zarindast A, Sharma A, Wood J (2021) Application of text mining in smart lighting literature—an analysis
of existing literature and a research agenda. Int J Inf Manag Data Insights 1(2):100032. https://​doi.​
org/​10.​1016/j.​jjimei.​2021.​100032. https://​linki​nghub.​elsev​ier.​com/​retri​eve/​pii/​S2667​09682​10002​52
Zhang Y, Xu B, Zhao T (2020) Convolutional multi-head self-attention on memory for aspect sentiment
classification. IEEE/CAA J Automatica Sinica 7(4):1038–1044. https://​doi.​org/​10.​1109/​JAS.​2020.​
10032​43. https://​ieeex​plore.​ieee.​org/​docum​ent/​91280​78/
Zhang W, Deng Y, Li X et al (2021a) Aspect sentiment quad prediction as paraphrase generation. In:
Moens MF, Huang X, Specia L et al (eds) Proceedings of the 2021 conference on empirical meth-
ods in natural language processing. Association for Computational Linguistics, Online and Punta
Cana, Dominican Republic, pp 9209–9219. https://​doi.​org/​10.​18653/​v1/​2021.​emnlp-​main.​726.
https://​aclan​tholo​gy.​org/​2021.​emnlp-​main.​726
Zhang W, Li X, Deng Y et al (2021b) Towards generative aspect-based sentiment analysis. In: Zong C,
Xia F, Li W et al (eds) Proceedings of the 59th annual meeting of the association for computa-
tional linguistics and the 11th international joint conference on natural language processing (vol-
ume 2: short papers). Association for Computational Linguistics, Online, pp 504–510. https://​doi.​
org/​10.​18653/​v1/​2021.​acl-​short.​64. https://​aclan​tholo​gy.​org/​2021.​acl-​short.​64
Zhang H, Chen Z, Chen B et al (2022) Complete quadruple extraction using a two-stage neural model
for aspect-based sentiment analysis. Neurocomputing 492:452–463. https://​doi.​org/​10.​1016/j.​neu-
com.​2022.​04.​027. https://​www.​scien​cedir​ect.​com/​scien​ce/​artic​le/​pii/​S0925​23122​20039​39
Zhang W, Li X, Deng Y et al (2022b) A survey on aspect-based sentiment analysis: tasks, methods, and
challenges. arXiv:​2203.​01054
Zhang W, Li X, Deng Y et al (2022) A survey on aspect-based sentiment analysis: tasks, methods, and
challenges. IEEE Trans on Knowl and Data Eng 35(11):11019–11038. https://​doi.​org/​10.​1109/​
TKDE.​2022.​32309​75
Zhang X, Xu J, Cai Y et al (2023) Detecting dependency-related sentiment features for aspect-level sen-
timent classification. IEEE Trans Affect Comput 14(1):196–210. https://​doi.​org/​10.​1109/​TAFFC.​
2021.​30632​59. https://​ieeex​plore.​ieee.​org/​docum​ent/​93689​87/
Zhang W, Zhang X, Cui S, et al (2024a) Adaptive data augmentation for aspect sentiment quad prediction.
In: ICASSP 2024—2024 IEEE international conference on acoustics, speech and signal processing
(ICASSP), pp 11176–11180. https://​doi.​org/​10.​1109/​ICASS​P48485.​2024.​10447​700
Zhang W, Zhang X, Cui S et al (2024b) Adaptive data augmentation for aspect sentiment quad prediction.
In: ICASSP 2024—2024 IEEE international conference on acoustics, speech and signal processing
(ICASSP), pp 11176–11180. https://​doi.​org/​10.​1109/​ICASS​P48485.​2024.​10447​700. https://​ieeex​
plore-​ieee-​org.​ezpro​xy.​auckl​and.​ac.​nz/​docum​ent/​10447​700
Zhao H, Yang M, Bai X et al (2024) A survey on multimodal aspect-based sentiment analysis. IEEE Access
12:12039–12052. https://​doi.​org/​10.​1109/​ACCESS.​2024.​33548​44. https://​ieeex​plore.​ieee.​org/​docum​
ent/​10401​113
Zhou J, Huang JX, Chen Q et al (2019) Deep learning for aspect-level sentiment classification: survey,
vision, and challenges. IEEE Access 7:78454–78483. https://​doi.​org/​10.​1109/​ACCESS.​2019.​29200​
75. https://​ieeex​plore.​ieee.​org/​docum​ent/​87263​53
A systematic review of aspect‑based sentiment analysis: domains,… Page 51 of 51 296

Zhou C, Wu Z, Song D et al (2024) Span-pair interaction and tagging for dialogue-level aspect-based senti-
ment quadruple analysis. In: Proceedings of the ACM on web conference 2024, WWW’24. Associa-
tion for Computing Machinery, New York, pp 3995–4005. https://​doi.​org/​10.​1145/​35893​34.​36453​55

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.