16 Machine Learning and Artificial Intelligence
J.-L. Kim (Ed.)
© 2023 The authors and IOS Press.
This article is published online with Open Access by IOS Press and distributed under the terms
of the Creative Commons Attribution Non-Commercial License 4.0 (CC BY-NC 4.0).
doi:10.3233/FAIA230763
Discovering Insights via Hybrid Thematic
Analysis: A Case Study on Disaster Risk
Reduction and Management for Legazpi
City, Albay
Mideth ABISADO a,1 Lany MACEDA b Ramon RODRIGUEZ a
Joseph Marvin IMPERIAL a Myron Darrel MONTEFALCON a Jay Rhald PADILLA a
Gizelle PONCE a
a College of Computing and Information Technologies,
National University, Sampaloc,
Manila 1008, Philippines
b Bicol University, Legazpi, Philippines
Abstract. Understanding the perceptions and experiences of a community regard-
ing disasters is crucial in effectively planning and implementing disaster strategies.
There are two known approaches to analyzing perceptions, qualitative and auto-
mated approaches for thematic analysis. This paper aims to investigate the strengths
and limitations of the mentioned approaches. Thus, using both approaches, this
study analyzed data from a focus group discussion about disaster risk reduction
and management and climate change adaptation conducted in a typhoon-prone city
in Legazpi City, Philippines. An inductive-deductive approach for the qualitative
method while a language model-assisted approach for the automatic method of
extracting prominent themes from the collected responses. The results show (dis)
similarities regarding themes obtained from the two approaches, specifically the
emphasis on concerns about the proper distribution of relief goods and donations,
proper early monitoring of potentially powerful typhoons, and other forms of threat,
including politically motivated ones. From these findings, we conclude the impor-
tance of incorporating a combined manual and automatic approach for thematic
analysis of natural language.
Keywords. Qualitative Approach, NLP-based Approach, Thematic Analysis,
DRMM-CCA, KeyBert Architecture
1. Introduction
The Philippines is highly vulnerable to natural disasters due to its location in the Pacific
Ring of Fire, making it prone to frequent typhoons and earthquakes [1]. According to
World Risk Index in the year 2022 [2], the Philippines has the highest disaster risk among
193 countries worldwide, scoring high in its exposure, vulnerability, susceptibility, lack
1 Corresponding Author: Mideth B. Abisado, College of Computing and Information Technologies, National
M. Abisado et al. / Discovering Insights via Hybrid Thematic Analysis: A Case Study on DRRM 17
of response capacities, and lack of adaptive capacities in the face of disasters. For the
Philippines to effectively respond to these disasters and mitigate their consequences, it
is crucial to understand the needs and perspectives of those affected. Listening to citi-
zens’ perceptions in Disaster Risk Reduction and Management (DRRM) can help ensure
that all perspectives and needs are considered in decision-making. The community has
first-hand experience and knowledge of local risks and hazards. Moreover, engaging the
citizens and getting their perspectives promotes inclusivity, improved planning, and in-
creases their sense of ownership of the community, which can provide valuable insights
into DRRM efforts.
The primary challenge for collecting and analyzing responses from the community
is that it is time-consuming, labor-intensive, and prone to human errors [3]. With this,
researchers have turned to use automated approaches via Natural Language Processing
(NLP). There have been several applications of NLP-based analysis in disaster-related
phenomena, including using NLP on social media analytics, as vast amounts of informa-
tion is generated in social networking platforms during disasters. The studies [4]-[5] im-
plemented NLP techniques to analyze disaster-related tweets from Twitter, understand-
ing behavior and obtaining valuable insights from the citizens in real-time. Similarly,
the work [6] performed sentiment analysis on disaster-related tweets, which enabled an
understanding of the public’s emotions and opinions. Aside from that, NLP-based ap-
proaches can be applied to cater to multilingual data [7] as information about disasters
may be shared in multiple languages, allowing for a more comprehensive understanding
of the event.
This paper aims to compare the effectiveness of the qualitative method and NLP-
based approaches in analyzing responses from a focus group discussion conducted in a
specific community about DRRM and CCA. Comparing these two approaches aims to
shed light on uncovering important and unique themes as a tool for accurate analysis.
2. Review of related literature
Thematic analysis is an approach that uncovers the underlying meanings and perspec-
tives that shape the experiences of individuals and communities. We divide the extensive
literature on thematic analysis approaches in two with respect to their methodology: (a)
conventional approaches using human interaction through interviews and focus group
discussions (FGD) and (b) automated approaches using Natural Language Processing.
2.1. Conventional thematic analysis with human interaction
Several papers [8,9] have applied a qualitative approach to analyzing data on disaster-
related topics in the Philippines. In the study of [8], the researchers examined the ex-
periences of youth survivors of Typhoon Haiyan through an interview to gather each
participant’s own narrative of the event. The study found that the experience of a nat-
ural disaster can profoundly impact the self and identity of youth survivors, leading to
changes in their physical and emotional well-being, relationships, and sense of self. Sim-
ilarly, the study of [9] investigated the evacuation response of affected communities dur-
ing the 2014 Mayon Volcano eruption in the Philippines. The authors used a qualitative
method through interviews with the evacuated residents, local officials, and first respon-
18 M. Abisado et al. / Discovering Insights via Hybrid Thematic Analysis: A Case Study on DRRM
ders to understand the factors that influenced the evacuation response and the challenges
communities faced during the evacuation. The study found that evacuation response was
influenced by community readiness, planning, communication, and transportation and
highlighted the importance of these factors for effective evacuation. However, the use of
the conventional qualitative approach for analysis also has some limitations as it can be
time-consuming and subjective which could lead to potential bias [11].
2.2. Thematic analysis using natural language processing techniques
Using NLP techniques for thematic analysis has become increasingly popular in recent
years. In the paper of [5], the researchers analyze the behavior of Filipino Twitter users
using Latent Dirichlet Allocation (LDA) and Principal Component Analysis to extract
the different topics discussed during a disaster. The study highlights that Filipino Twit-
ter users shared various information related to the disaster, including updates on the dis-
aster’s progression, personal experiences, and calls for help. Similarly, in the works of
[12] and [13], the authors performed NLP-based techniques like topic modeling and sen-
timent analysis on typhoons and earthquake-related tweets in the Philippines. Another
NLP-based paper [14], utilizes 976 suggestions on how their village can help them better
prepare for a disaster implemented computational methods, specifically topic modeling
and word2vec, to assist in the analysis of qualitative data on disaster risk reduction sug-
gestions. Based on the results, computational methods improved the efficiency and ac-
curacy of the data analysis process. The study’s findings show that the participants give
importance to community preparedness for an emergency, helping the village in clean-up
drives, and awareness through seminars and information dissemination.
2.3. Hybrid approaches
Combining the conventional qualitative thematic analysis with Natural Language Pro-
cessing (NLP) techniques offer several advantages that merit exploration. By combin-
ing both approaches, researchers can leverage the strengths of each method to enhance
the accuracy and efficiency of their analysis. Moreover, NLP techniques can help re-
searchers identify themes and patterns that might not have been apparent in manual anal-
ysis, thus leading to more comprehensive and nuanced insights. Ultimately, integrat-
ing conventional and NLP-based approaches can provide a complete understanding of
the data and facilitate the development of evidence-based strategies and interventions in
disaster-related events.
2.4. Data collection through focus group discussions
A Focus Group Discussion – Structured Learning Exercise (FGD-SLE) was em-
ployed as a collection method as it provides a way to collect rich in-depth data from a
group of people. Following the Risk-Based Approach in Community-Based DRRM, we
have decided to utilize FGD-SLE to uncover new insights regarding the local grassroots
community’s knowledge and understanding of a complex issue such as DRRM-CCA.
This case study gathered DRRM-CCA perceptions and responses in the community
of Brgy. Padang, Legazpi City, Albay in the Philippines. The chosen community is
based on the endorsement of the Local Government Unit (LGU) as it is labeled as village
at risk and the mentioned community has become infamous for the onslaught of Typhoon
M. Abisado et al. / Discovering Insights via Hybrid Thematic Analysis: A Case Study on DRRM 19
Reming in the year 2006 [15] which caused a lot of casualties. Additionally, based on
the Risk-based approach assessment, Brgy. Padang is both a high vulnerability index and
capacity or response measures. Thus, this community is selected for the initial FGD-SLE
to gather DRRM-CCA responses.
In the data collection process, it employed a purposive sampling method for the
FGD-SLE. Wherein the participants of the FGD-SLE included sectoral and marginalized
groups from Brgy. Padang, Legazpi City, Albay. A total of 16 participants included a
diverse group of individuals such as women, youth, farmers, fisherfolks, persons with
disability, the elderly, and local leaders of Brgy. Padang age ranges from 24 years old to
70 years old. The FGD-SLE was facilitated in a closed-door venue with activities like
group workshops, group reporting, and breakout sessions. During the FGD-SLE, the re-
sponses were recorded using an audio recorder, with a signed waiver of data privacy and
protection for each participant. A total of 4 hours of audio recording was collected. After
the FGD-SLE, the audio data was validated, and the recording was transcribed before
undergoing the analysis.
3. Experiment setup
This paper aims to provide an analysis comparing the qualitative thematic analysis ap-
proach with the NLP-based technique in getting themes and insights from the FGD-SLE
data collected from the participants. This will help understand the strengths and limita-
tions of both approaches empirically. The goal of the thematic analysis is to identify and
understand the themes, patterns, or structures present throughout the FGD-SLE data that
would help develop the questions for the DRRM-CCA component of the project.
3.1. Inductive-deductive approach
For the conventional qualitative approach, an inductive-deductive analysis was utilized,
wherein the inductive phase was used to reveal the central themes. An inductive research
method involves an analysis of texts and reasoning that stream from a set of empirical ob-
servations toward a systematic formulation of theories or conclusions [16]. This method
has strengths drawn for qualitative research, whereby flexibility and the bottom-up, it-
erative process of discovering themes and patterns from raw data take priority over pre-
determined ideas and notions. We have decided to use an inductive approach as the study
entails a more progressive qualitative, community-based, and participatory approach to
research. Afterward, a deductive analysis was employed, to derive codes from the central
themes of the specified community.
3.2. Language-model assisted keyword extraction
For the automated extraction of themes from the response data, we use KeyBERT [17],
a simple neural language model-based architecture for keyword extraction from doc-
uments. KeyBERT leverages the use of text representations from a Bidirectional En-
coder Representations from Transformers (BERT) language model called embed-
dings [18]. These embeddings contain rich linguistic information about the text such as
semantic knowledge. We use this approach over statistical-based or static keyword ex-
traction methods as it is a more applicable study since our goal is to obtain contextually-
20 M. Abisado et al. / Discovering Insights via Hybrid Thematic Analysis: A Case Study on DRRM
similar terms that can represent the entire document. The process of KeyBERT keyword
extraction starts with transforming a dataset D = d1 , d2 , d3 ...dn into its corresponding
document embeddings H(d) by feeding it to a selected BERT model. Then, word-level
representations H(w) are also extracted for all unique terms of the document. Finally,
each of the terms from the extracted vocabulary will be paired with the document via co-
sine similarity COS(di , wi ), where the top words garnering the highest similarity scores
are labeled as the most prominent keywords that can represent the context of the dataset.
For this study, since we are dealing with responses from the participants in tran-
scribed in Tagalog, we use the robust version Tagalog BERT model called RoBERTa as
the main language model of choice [19]. We set the parameters of KeyBERT to generate
10 potential keyword groups with each containing 5 candidate keywords. We then eval-
uate these groups to obtain the top five most semantically related to the document for
analysis.
3.3. Evaluation metric
For any automatic analysis method conducted with real-world data, it is important to
evaluate the results obtained from the experimental processes to ensure the reliability of
the findings. In this work, we evaluate each of the 10 keyword groups generated from
the KeyBERT experiment using the coherence-type metric UMASS [20]. The UMASS
score is an intrinsic document co-occurrence based metric where (wi , w j ) denotes the
number of times both words wi and w j occurred in a document D normalized by the
number of times wi appears in the document. The formula is listed below:
D(wi , w j ) + 1
UMASS(wi , w j ) = log (1)
D(wi )
4. Prominent disaster-related Themes
4.1. Themes from focus group discussions
Based on processing the maps and the group reports, significant themes were decoded
from the FGD-SLE activity. Table 1 presents the main themes including Perceptions of
DRRM-CCA, LGU’s Programs and Projects, and Community-Based Solutions and
Recommendations.
From the first theme, the participants shared personal reflections, indigenous knowl-
edge, and perceived problems with DRRM-CCA in their community. From the reported
risk-mapping activity, it is evident that the participants view typhoons as the most sig-
nificant risk event, followed by volcanic eruptions at Mt. Mayon, the rainy season as
medium risk, and earthquakes as low risk. Results from the activity suggest that the
participants understand the local disaster risks in their village and that typhoons are
perceived as the most significant hazard in their area. The second code highlights the
perceived DRRM-CCA-related challenges, including trauma, financial difficulty, water
scarcity, lack of sanitary toilets, stress, diseases, difficulty with household chores, and
various health problems. Additionally, the participants reported a lack of proper knowl-
edge and preparation for disasters, with only a basic understanding of the ”drop, cover,
and hold” method—a prominent must-do in the government’s earthquake information
M. Abisado et al. / Discovering Insights via Hybrid Thematic Analysis: A Case Study on DRRM 21
and education campaigns. The vulnerability map covers this theme, as the participants
also reported the importance of social assistance and social safety nets (e.g., cash or
crops, water and food resources, and medical support) during disaster recovery and re-
habilitation.
The second theme relates to the FGD-SLE activity gauging the participants’ under-
standing of the LGU’s programs, projects, and policies in the community. It is evident
that the community participants are highly aware of the LGU’s programs and policies
on DRRM-CCA. The participants can discuss scenarios and hazards where the LGU
has prepared evacuation centers and early-warning equipment, divided responsibilities
among staff, assisted senior citizens and traumatized evacuees, monitored health and
food, and provided relief assistance to the affected constituents of Brgy. Padang. The
method also captures the strengths and assets of the community through their positive
perceptions of their LGU’s programs and policies, including the character of leaders,
tools, local expertise, leadership, and connections to other groups. According to the par-
ticipants, the emergency local meeting and village staff play a crucial role in applying
these strengths and ensuring the community’s preparedness for disasters.
The last theme focuses on community-based solutions and disaster preparedness
and resilience recommendations. This theme is encapsulated in the last group’s recom-
mendatory report on the community’s actions to build resilience against disasters. The
participants emphasized in their report the importance of individuals being self-sufficient
and not overly dependent on the village LGU. To enhance resilience against disasters
and climate change, the participants suggested educating the youth and increasing their
early awareness and orientation to the impacts of disasters. They believe that confidence,
courage, and faith in God are vital for the community to handle disasters effectively. Ac-
cording to them, with or without LGU support, the local people of Brgy. Padang must
enhance their self-sufficiency and empower themselves to handle the various disasters
in their way. They further noted that this is the only way for them to be sustainable in
disaster prevention.
Table 1. Results from qualitative thematic analysis approach: inductive-deductive.
Theme Code Examples
The participants view typhoons as high risk,
Personal Reflections and Knowledge on DRRM-CCA followed by a volcanic eruption at Mt. Mayon.
Perceptions on DRMM-CCA
They view the rainy season as medium risk and earthquakes as low risk.
Challenges reported by participants include trauma,
financial difficulties, water scarcity,
Perceived Problems on DRRM-CCA in the Community
lack of sanitary toilets, stress, and various illnesses
such as colds, coughs, and fevers.
The community exhibits strengths that enhance
its crisis resilience, such as discussing scenarios, preparing evacuation
centers and equipment, aiding vulnerable members,
Awareness of LGU Programs and monitoring health and food, and providing relief assistance.
LGU DRMM-CCA Programs
Policies on DRRM-CCA in the Community The emergency CDRRMO local meeting and
and Policies
barangay staff help identify needs and damages,
prepare centers and equipment, monitor the situation, check supplies,
and distribute aid.
Participants in Brgy. Padang reported ample disaster
preparation programs from the LGU,
Perception of Effectiveness of
including having food and non-food items
LGU DRRM-CCA Programs and Policies
as essential goods, health, and security protocols,
and resources like medicine, flashlight, and gas.
Participants in Brgy. Padang reported sufficient
disaster supplies, including food, medicine,
Community-based Solutions
Community Initiatives and Recommended Actions flashlights, and diesel. They emphasized
and Recommendations
the need for self-sufficiency, youth education, and
faith in God to increase resilience and effective disaster response.
22 M. Abisado et al. / Discovering Insights via Hybrid Thematic Analysis: A Case Study on DRRM
Table 2. Results from performing automatic keyword extraction using the KeyBERT architecture and Tagalog
RoBERTa model.
Labelled Topics Groups Keywords UMASS
General Disaster Terms* vulnerabilities, disasters, climate, point, cause -0.207
Quarrying Concerns* nagqquarry, disaster, pagpuputol, behavior, basura -0.769
Description of Citizens kabataang, marites, tambay, nakapagtapos, gurang -14.072
Advisory Concerns late, political, advisory, suspensions, advisories -3.033
Vocational Education beneficial, tesda, nakakatulong, kakatapos, kaalaman -7.186
Suspension Concerns nagbaha, nagsuspinde, klase, happy, bahay -9.235
Preparation and Monitoring* pagprepara, monitoring, typhoon, centers, bhw -0.069
Donations* mabibigay, diskarte, kailangan, pangangailangan, laging -0.496
Volcanic Eruptions forced, mayon, bulkan, evacuation, ashfall -9.787
Local Politics* nadadamay, barilan, pulitiko, lugar, violence 0.009
4.2. Themes from automatic Analysis
Table 2 presents the top 10 extracted thematic keywords using the KeyBERT method
with the Tagalog RoBERTa model. From the results, we highlight five topics which can
be considered prominent due to their relatively high UMASS score compared to the oth-
ers. These topics are in the context of general disaster terms, concerns on quarrying,
preparation and monitoring, concerns on donations, and local politics. Similar to re-
lated work [14], these themes were manually labeled based on the context provided by
their supporting topic keywords as seen in the second column of Table 2.
From the prominent topics, we draw similarities and dissimilarities with the con-
ventional qualitative approach presented in Table1. The common theme that appeared
from both methods can be attributed to the context of concerns on donations. Asso-
ciated words that make up this topic through the KeyBERT approach are kailangan or
need, magbibigay or to give, and pangangailangan or needs. These words are contex-
tually usable during times of need as seen in the second theme from the conventional
qualitative approach which is LGU DRRM-CCA Programs and Policies. Interestingly,
one word from the automatic method, diskarte or the act of hustling or resourceful, may
be associated with the special attitude of Filipino people to exhaust all means to survive
in times of disasters.
Another common theme from both methods is preparation and monitoring. As
mandated, ensuring enough preparation of food stocks and accurate monitoring of any
signals of incoming natural disasters such as heavy rains are under the responsibility
of LGU. This concern is also highlighted in the second theme from the qualitative ap-
proach. Associated words formed through the KeyBERT method are pagprepara or act
of preparation, centers, monitoring, and typhoon. Another prominent theme is associ-
ated with quarrying concerns, which can be linked to concerns of local communities
towards illegal quarrying or mining activities and have been a long-time problem in the
city2 . Supporting thematic words building this context are nagququarry or quarrying,
disaster, pagpuputol or cutting, behavior, and basura or trash. This particular concern
can also be seen in the results from the qualitative approach which can be classified un-
der the second code of the first theme which is Perceived Problems with DRRM-CCA in
the Community.
2 https://www.pna.gov.ph/articles/1122671
M. Abisado et al. / Discovering Insights via Hybrid Thematic Analysis: A Case Study on DRRM 23
Lastly, using the automatic approach extracted an interesting topic which is local
politics. This particular theme is associated with convincingly negative terms such as
nadadamay or affected, barilan or shooting, and violence. This theme may suggest that
even if there are other major concerns regarding natural disasters in the community, other
threats to safety and security, including those that are politically motatived, can still cause
worry and concerns to members of the community.
5. Conclusions and future Work
This study explored the perceptions and experiences of a typhoon-prone community
in Legazpi City, Philippines, on disaster risk reduction and management (DRRM) and
climate change adaptation (CCA). Our findings demonstrate the importance of under-
standing community perspectives to develop effective disaster management strategies.
We used the conventional qualitative and NLP-based approaches to analyze the collected
responses and identified key themes from both approaches. Comparing the results of
the two approaches, we found strong similarities; both emphasized the need for proper
distribution of relief goods and early monitoring of potentially powerful typhoons and
other forms of threats, including politically motivated ones. This highlights the capability
of using an NLP-based approach in data analysis as it can obtain similar insights from
the conventional qualitative approach while uncovering underlying themes. On the other
hand, some themes were only present in one of the approaches, indicating the impor-
tance of implementing both approaches, as it can provide a clearer understanding of the
responses of the community.
Acknowledgments
We gratefully acknowledge the financial support the Philippine Commission on Higher
Education (CHED) provided under Grant No. 2021-007. This funding has been instru-
mental in advancing our research endeavors and enabling us to make significant contribu-
tions to our field. We extend our sincere appreciation for CHED’s continued commitment
to the advancement of scientific knowledge and innovation.
References
[1] Bollettino V, Alcayna T, Enriquez K, Vinck P. Perceptions of disaster resilience and preparedness in the
Philippines. Harvard University: Cambridge, MA, USA. 2018 Jun.
[2] Franziska A, Lotte K, Beáta P, Kristin BS, Sören S, Daniel W, Bündnis EH. World Risk Report
2022—Focus: Digitalization.
[3] Mason H, Patil DJ. Data Driven. Creating a data culture; 2015.
[4] Takahashi B, Tandoc Jr EC, Carmichael C. Communicating on Twitter during a disaster: An analysis of
tweets during Typhoon Haiyan in the Philippines. Computers in human behavior. 2015 Sep 1;50:392-8.
[5] Lee Boaz J, Ybañez M, De Leon MM, Estuar MR. Understanding the behavior of Filipino Twitter users
during disaster. GSTF Journal on Computing (JoC). 2013 Aug;3:1-6.
[6] Zhang T, Cheng C. Temporal and spatial evolution and influencing factors of public sentiment in natural
disasters—a case study of typhoon haiyan. ISPRS International Journal of Geo-Information. 2021 May
5;10(5):299.
24 M. Abisado et al. / Discovering Insights via Hybrid Thematic Analysis: A Case Study on DRRM
[7] Fuadvy MJ, Ibrahim R. Multilingual sentiment analysis on social media disaster data. In2019 Interna-
tional Conference on Electrical, Electronics and Information Engineering (ICEEIE) 2019 Oct 3 (Vol. 6,
pp. 269-272). IEEE.
[8] Delante N. Natural disasters reconstruct the self: A thematic analysis of youth survivors’ narratives about
typhoon haiyan. Journal of Narrative and Language Studies. 2019 Dec 30;7(13):113-29.
[9] Martinez-Villegas MM, Solidum RU, Saludadez JA, Pidlaoan AC, Lamela RC. Moving for safety: a
qualitative analysis of affected communities’ evacuation response during the 2014 Mayon Volcano erup-
tion. Journal of Applied Volcanology. 2021 Dec;10(1):1-24.
[10] Neuendorf KA. The content analysis guidebook. sage; 2017.
[11] Braun V, Clarke V. Using thematic analysis in psychology. Qualitative research in psychology. 2006 Jan
1;3(2):77-101.
[12] Maceda LL, Llovido JL, Palaoag TD. Corpus analysis of earthquake related tweets through topic mod-
elling. International Journal of Machine Learning and Computing. 2017 Dec;7(6):194-7.
[13] Imperial JM, Orosco J, Mazo SM, Maceda L. Sentiment analysis of typhoon related tweets using stan-
dard and bidirectional recurrent neural networks. arXiv preprint arXiv:1908.01765. 2019 Aug 3.
[14] Gorro K, Ancheta JR, Capao K, Oco N, Roxas RE, Sabellano MJ, Nonnecke B, Mohanty S, Crittenden
C, Goldberg K. Qualitative data analysis of disaster risk reduction suggestions assisted by topic modeling
and word2vec. In2017 International Conference on Asian Language Processing (IALP) 2017 Dec 5 (pp.
293-297). IEEE.
[15] Orense RP, Ikeda M. Damage caused by typhoon-induced lahar flows from Mayon Volcano, Philippines.
Soils and foundations. 2007 Dec 1;47(6):1123-32.
[16] Vanderstoep SW, Johnson DD. Research methods for everyday life: Blending qualitative and quantitative
approaches. John Wiley & Sons; 2008 Dec 22.
[17] Grootendorst M. KeyBERT: Minimal keyword extraction with BERT. Zenodo. 2020.
[18] Devlin J, Chang MW, Lee K, Toutanova K. Bert: Pre-training of deep bidirectional transformers for
language understanding. arXiv preprint arXiv:1810.04805. 2018 Oct 11.
[19] Cruz JC, Cheng C. Improving large-scale language models and resources for filipino. arXiv preprint
arXiv:2111.06053. 2021 Nov 11.
[20] Mimno D, Wallach H, Talley E, Leenders M, McCallum A. Optimizing semantic coherence in topic
models. InProceedings of the 2011 conference on empirical methods in natural language processing
2011 Jul (pp. 262-272).
