International Journal of Computer Applications (0975 – 8887)

Volume 186 – No.46, November 2024

Multi-Modal Machine Learning for Political Video

Advertisement Analysis: Integrating Audio, Textual,
and Visual Features
Moulik Kumar Satish Gopalani Pranav Gupta
Indian Institute of Technology, Kharagpur PubMatic Inc PubMatic Inc

ABSTRACT ad targeting, and enhance user experience. As digital ad spending

continues to grow, reaching $455.3 billion globally in 2021 and
This paper presents a novel framework for the automated classi- surpassing $1 trillion by 2025, the need for sophisticated content
fication, tagging, and issue-level sentiment analysis of video ad- analysis tools has never been greater [2].
vertisements using advanced machine-learning techniques. The Political advertising, particularly in video format, represents a sig-
proposed multi-pass approach leverages audio transcription, Op- nificant and growing segment of digital advertising. In the United
tical Character Recognition (OCR), and video feature extrac- States, political ad spending is projected to reach $12.3 billion in
tion to achieve high accuracy in distinguishing between political 2024, a 30% increase from $9.6 billion in 2020. Digital media is
and non-political content. The research introduces robust meth- ods expected to account for approximately $3.5 billion of this spend-
for candidate identification for political videos using phrase ing, nearly tripling from the previous election cycle [3]. This surge
matching and fuzzy logic, as well as issue tagging and senti- ment in digital political advertising necessitates advanced classification,
analysis utilizing natural language processing algorithms. The verification, and analysis systems.
system demonstrates significant improvements over existing This research focuses on developing a robust framework for classi-
methods, achieving 99.2% accuracy in political ad classification fying political video ads, verifying candidates, and analyzing issue-
when combining audio and OCR data. Furthermore, the devel- level sentiment using advanced text processing and machine learn-
oped issue-level sentiment analysis provides granular insights into ing techniques. The study addresses not only the binary classifi-
the emotional tone of political messaging. This research con- cation of political versus non-political content but also the more
tributes to the growing field of content moderation in digital ad- nuanced tasks of candidate/party identification, political issue tag-
vertising, offering valuable insights for publishers, researchers, and ging, and sentiment detection at the issue level. This comprehen-
policymakers in the realm of political communication. sive approach aims to provide publishers with granular control over
their ad inventory, enabling them to curate content that aligns with
General Terms their brand values and audience expectations.
Political advertising, Video classification, Machine learning, Natural lan- By implementing such a system, publishers can maintain platform
guage processing, Content moderation integrity, ensure regulatory compliance, and create a more transpar-
ent and effective digital advertising ecosystem. This research con-
tributes to the growing body of work on content moderation in the
Keywords digital age, with a specific focus on the unique challenges posed by
Political advertising, Video classification, Machine learning, Nat- political video advertising.
ural language processing, Content moderation, Candidate verifica-
tion, Sentiment analysis
2. CURRENT APPROACHES
1. INTRODUCTION
The classification and tagging of political video advertisements
Video classification techniques have evolved rapidly in recent have traditionally been performed manually, a process that is both
years, leveraging advances in machine learning and computer vi- time-consuming and prone to human error. The rapid growth in
sion to automatically categorize visual content. These methods, political advertising, particularly on digital platforms, has neces-
which analyze features from video frames, audio, and metadata, sitated the development of automated methods to handle the sheer
find applications across various domains, including content moder- volume of content effectively.
ation, recommendation systems, and advertising [1]. Recent research has focused on using deep learning techniques for
In the digital advertising landscape, video classification has become video classification. Dhakal (2019) [4] investigated various deep-
increasingly crucial. The Advertisement Technology (AdTech) in- learning methods for classifying political video advertisements by
dustry relies on these techniques to ensure brand safety, improve combining text features with video classification techniques. This

49
 International Journal of Computer Applications (0975 – 8887)
Volume 186 – No.46, November 2024
—The process extracts and analyzes text visible in the video
study highlights the potential of deep learning to significantly im- frames, particularly focusing on the final frames where dis-
prove the accuracy of political ad classification claimers often appear.
Another study by Belhaouari et al. (2023) [5] provides a com- —This additional textual information helps to catch political
prehensive review of deep learning models for video classifica- ads that might not have clear audio indicators.
tion, emphasizing the importance of network architecture, evalua-
(3) Third Pass: Comprehensive Audio, OCR, and Video Fea-
tion criteria, and benchmark datasets. This review indicates that ad-
ture Classification
vanced deep learning methods, such as Convolutional Neural Net-
works (CNNs) and transformers, have made substantial progress in —In the final pass, the system incorporates visual features ex-
the automatic classification of videos. tracted from the video frames.
Grigsby and Fowler (2020) explored the political ideology of ads —This comprehensive analysis combines all available data:
by analyzing image and text content, focusing on the type of ad audio transcription, OCR text, and video features.
sponsor and the ideological leanings of the ad. Their work under- —This pass is designed to make final decisions on ads that
scores the importance of accurately classifying political ads to un- were ambiguous in the previous stages.
derstand digital campaigning’s characteristics and its impact on po-
litical communication [6]. 3.3 System Architecture
Despite these advancements, the task of political ad classification Figure 1 illustrates the architecture of the 3-pass classification sys-
still faces challenges due to the nuanced and varied nature of politi- tem. The process flow can be summarized as follows:
cal content. Current approaches often require large labelled datasets
for training, which may not always be available. Moreover, the abil- (1) Input video advertisements are processed to extract audio tran-
ity to handle multi-modal data—combining text, audio, and visual scriptions.
features—remains a critical area for improvement. (2) Initial classification is performed using audio features.
(3) For ambiguous cases, OCR is applied to extract on-screen text,
3. PROPOSED SOLUTION and classification is refined.
3.1 Problem Overview (4) In the final pass, video features are incorporated for compre-
hensive classification.
This research aims to develop a comprehensive system for analyz-
ing political video advertisements. The main objectives of the sys- (5) The output is a binary classification of each ad as either politi-
tem are to: cal or non-political.

(1) Distinguish between political and non-political video ads 3.4 Subsequent Analysis
(2) Identify candidates mentioned in political ads
Following the classification of an advertisement as political, the
(3) Detect the political issues discussed in the ads system performs additional analyses, including:
(4) Analyze the sentiment associated with each issue and the over-
all ad tone —Candidate Identification: Detecting and identifying political
candidates mentioned in the ad.
These objectives present a complex challenge that requires analyz-
ing multiple aspects of video content, including spoken words, on- —Issue Detection: Identifying the political issues discussed in the
screen text, and visual elements. ad.
—Sentiment Analysis: Analyzing the sentiment associated with
3.2 Framework Approach each detected issue and the overall ad tone.
To address the challenge of political ad classification, this paper These subsequent steps, while crucial for comprehensive political
proposes a multi-pass framework that leverages different types of ad analysis, are separate from the initial classification framework
data extracted from video advertisements. This approach is based and will be discussed in detail in later sections of this paper.
on analysis of political ad characteristics and regulatory require-
ments. The dataset indicates that approximately 90% of political 4. DATA PREPARATION
video ads in the U.S. contain spoken political commentary, making
audio analysis an effective first pass. The second pass incorporates This study utilizes a dataset of 2,071 video creatives obtained from
OCR to capture text-based disclaimers, which are mandated for po- a leading adtech company, comprising 1,082 non-political and 989
litical ads [7] and present in 95% of the samples. The approach political advertisements. This dataset was carefully curated to rep-
consists of three main stages, each incorporating additional data to resent the diverse range of content typically encountered by pub-
refine the classification: lishers in real-world scenarios.
The video creatives exhibit significant variety in terms of content,
(1) First Pass: Audio-based Classification duration, production quality, target audience, and advertising objec-
—In this initial stage, the system transcribes the audio content tives. This diversity ensures that the model is trained on a represen-
of the video advertisement. tative sample of advertisements, enhancing its ability to generalize
—The analysis processes the transcribed text to make a pre- to new, unseen content.
liminary classification. By maintaining a near-balanced distribution between political and
—This pass is computationally efficient and allows for quick non-political ads, the methodology minimizes potential biases in
filtering of clearly non-political content. model training and evaluation. This approach enables the system to
(2) Second Pass: Audio and OCR-based Classification develop a nuanced understanding of the subtle distinctions between
—For ads that aren’t accurately classified in the first pass, the political and non-political content, particularly in cases where ads
system incorporates Optical Character Recognition (OCR) may share similar language or imagery.
data.

50
 International Journal of Computer Applications (0975 – 8887)
Volume 186 – No.46, November 2024

Fig. 1: 3-Pass Classification System Architecture

All creatives were anonymized and processed in compliance with 5.2 Optical Character Recognition (OCR)
relevant data protection regulations and ethical guidelines for re-
search. For extracting text from video frames, the study employed an ac-
cessible OCR engine [11] known for its effectiveness in detecting
and recognizing text in various orientations and styles. This tool is
5. METHODOLOGY particularly useful for political ads, which often include essential
5.1 Audio Transcription and Analysis on-screen text such as disclaimers and candidate names in the final
frames [7].
The audio files were transcribed using OpenAI’s Whisper model The OCR process involves:
[8], an open-source, state-of-the-art automatic speech recognition
(ASR) system. Whisper is renowned for its robust performance (1) Frame Extraction: Selecting key frames from the video, with
across diverse accents, dialects, and noisy environments, making it a focus on the final frames where political disclaimers often
ideal for the analysis of political advertisements where clarity and appear [12].
accuracy of transcription are paramount. (2) Text Detection: Identifying regions in the frames that contain
Key steps in the audio analysis include: text.
(3) Text Recognition: Converting the detected text regions into
(1) Transcription: Converting spoken words to text using ASR. machine-readable text.
(2) Text Preprocessing: Cleaning and normalizing the transcribed (4) Text Processing: Applying similar preprocessing steps as used
text, including removing punctuation, converting to lowercase, for the audio transcription.
and handling common abbreviations [9].
(3) Feature Extraction: Applying TF-IDF (Term Frequency- 5.3 Video Feature Extraction
Inverse Document Frequency) vectorization to convert the text
into a numerical format suitable for machine learning algo- In the third pass, the system extracts visual features directly from
rithms [10]. the video content, focusing on the last 7.5 seconds of each video.
This choice is motivated by the common practice in political adver-
tisements to display crucial information towards the end of the ad
[7]. The key features extracted include:
TF-IDF(t, d, D) = TF(t, d) × IDF(t, D)
(1) Color Analysis:
—Percentage of red pixels
Number of times term t appears in document d —Percentage of blue pixels

TF(t, d) = —Combined percentage of red and blue pixels

Total number of terms in document d (2) Face Detection:
—Total number of faces detected
,N , —Average number of faces per frame
IDF(t, D) = log
—Average number of faces per second
nt
—Maximum number of faces in any single frame
—Face density (faces per second)
N = Total number of documents in the corpus (3) Video Duration:
nt = Number of documents containing the term t —Total length of the video in seconds

51
 International Journal of Computer Applications (0975 – 8887)
Volume 186 – No.46, November 2024
6.2 Classification Models
The study experiments with several machine learning algorithms
for the binary classification task of distinguishing between political
and non-political ads. The models evaluated include:
—Support Vector Machines (SVM): Effective for high-
dimensional spaces but may struggle with large datasets.
—Logistic Regression: Offers high explainability through accessi-
ble regression coefficients for TF-IDF features, and is efficient,
but may underperform with complex non-linear relationships.
—Random Forest: Handles non-linear relationships well and is less
prone to overfitting, but can be computationally intensive.
—Gradient Boosting Machines: Often achieves high accuracy but
may be prone to overfitting without careful tuning.
Fig. 2: Example Political Advertisement [13] —Neural Networks: Capable of capturing complex patterns but re-
Red-Blue %: 21.51% quires substantial data and computational resources.
Red %: 7.82% Each model is trained and evaluated using stratified k-fold cross-
Blue %: 13.69% validation [17] to ensure robust performance estimates. The results
show that Logistic Regression offered the best balance between
The study utilizes OpenCV [14] for video processing and the Haar performance and explainability. The ability to interpret the regres-
Cascade classifier [15] for face detection, allowing for efficient pro- sion coefficients for TF-IDF features in Logistic Regression pro-
cessing of large numbers of video files while capturing key visual vides valuable insights into the most influential terms for political
characteristics that may indicate political content. ad classification.

6.3 Evaluation Metrics

To assess the performance of the study’s models, the following
evaluation metrics were used:
(1) Accuracy:
TP + TN
Accuracy = (1)
TP + TN + FP + FN
(2) Precision:
TP
Precision = (2)
TP + FP

(3) Recall (also known as Sensitivity or True Positive Rate):

TP
Recall = (3)
TP + FN

(4) F1-Score:
Fig. 3: Haar Cascade Algorithm [16]
Precision × Recall
F1-Score = 2 × (4)
Precision + Recall

(5) Area Under the Receiver Operating Characteristic Curve

(AUC-ROC):
6. MODEL TRAINING AND EVALUATION
∫ 1
6.1 Feature Engineering AUC-ROC = TPR(T ) × FPR ′ (T ) dT (5)
0

After extracting features from audio transcriptions, OCR text, and

Where TPR is the True Positive Rate and FPR is the False Pos-
video content, the study combines these multi-modal features into
itive Rate at threshold T.
a unified representation for each advertisement. The feature set in-
cludes: Where:
(1) TF-IDF vectors from audio transcriptions —TP = True Positives
(2) TF-IDF vectors from OCR-extracted text —TN = True Negatives
(3) Video features (color analysis, face detection metrics, and du- —FP = False Positives
ration) —FN = False Negatives

52
 International Journal of Computer Applications (0975 – 8887)
Volume 186 – No.46, November 2024
A final ”match score” was utilized by proportionately weighing the
6.4 Classification Model Results matching scores of the first, last, and compound names:
The multi-pass approach demonstrated significant improvements
over single-modality classification methods. Table 1 presents the Match Score = 0.3 · first phonetic + 0.3 · last phonetic
performance metrics for the best-performing model at each stage

of the classification process. +0.1 · first fuzzy + 0.1 · last fuzzy (8)
The results show that incorporating OCR and video features pro- +0.1 · full name + 0.1 · jaro
gressively improves the classification performance, with the final
model achieving 99.5% accuracy. 7.4 Verification Against Official Database

Detected candidate names are verified against the Federal Election

6.5 Feature Importance Analysis
Commission (FEC) candidate database [24]. This step ensures that
To understand the contribution of different features to the classifi- the identified names correspond to actual political candidates, re-
cation decision, the study analyzed the feature importance scores. ducing false positives.
Tables 2 and 3 show the top 10 most significant TF-IDF features
from audio transcriptions and OCR-extracted text, respectively. 8. ISSUE DETECTION AND SENTIMENT
These tables highlight the most influential terms in determining
ANALYSIS
whether an advertisement is political.
8.1 Issue Detection
7. CANDIDATE IDENTIFICATION AND The proposed system employs a multi-faceted approach to detect
VERIFICATION political issues discussed in the advertisements:
Following the classification of an ad as political, the proposed sys- —Predefined Issue Dictionary
tem performs candidate identification and verification. This process The study maintain a comprehensive dictionary of politi- cal
involves several steps: issues (e.g., ”healthcare”, ”immigration”, ”environmen- tal
politics”), each associated with relevant keywords and phrases.
7.1 Key Phrase Extraction
—Machine Learning Classification
The study defines a set of key political phrases, each associated A multi-output classifier (MultiOutputClassifier with Lin-
with a weight reflecting its importance in identifying candidates. earSVC) [25] is trained on a labeled dataset to predict the pres-
These phrases are matched within the ad text using regular expres- ence of predefined issues in the ad text.
sions [18], with specific instructions on where to look for candidate —Topic Modeling
names relative to the phrase (before, after, or both directions) and The study uses Latent Dirichlet Allocation (LDA) [26] to detect
within a specified character limit. potential new or emerging issues not covered by the predefined
set. This allows the system to adapt to evolving political dis-
7.2 Named Entity Recognition course.
The study utilized Named Entity Recognition (NER) [19] to iden-
tify potential candidate names within the text. This step helps in 8.2 Sentiment Analysis
capturing names that might not be directly associated with the pre- For each detected issue, as well as for the overall ad, the study
defined key phrases. performs sentiment analysis using a multi-faceted approach:
7.3 Fuzzy Matching —Issue-Specific Sentiment
Each predefined issue has associated lists of positive and nega-
To account for variations in name spelling and common nicknames,
tive phrases. The study counts occurrences of these phrases to
the study employs fuzzy matching techniques [20]. This includes:
determine the sentiment expressed towards each specific issue.
(1) Levenshtein Distance [21]: —General Sentiment
max (i, j) if min(i, j) = 0,

leva,b(i − 1, j) + 1 The study employs a pre-trained sentiment analysis model (Dis-

leva,b (i, j) =
min leva,b(i, j − 1) + 1 otherwise.
(6) tilBERT fine-tuned on SST-2) [27] in combination with TextBlob
leva,b(i − 1, j − 1) + [ai ̸= bj ] [28] to assess the overall sentiment of the text.

Where [ai ̸= bj] is 1 if ai ̸= bj, and 0 otherwise. —Context-Aware Adjustment

(2) Jaro-Winkler Similarity [22]: The sentiment scores are adjusted based on the presence of nega-
tion words (e.g., ”not”, ”never”) and other contextual cues.
JW (s1, s2) = J(s1, s2) + lp(1 − J(s1, s2)) (7)
Where J is Jaro Similarity, l is common prefix length (max 4), 9. CONCLUSION AND FUTURE WORK
p is scaling factor (usually 0.1). This paper presents a comprehensive framework for the auto-
(3) Metaphone algorithm [23]: Applies phonetic rules to convert mated classification, candidate verification, and issue-level sen-
names to codes. Key rules: timent analysis of political video advertisements. The proposed
—Drop duplicate adjacent letters, except C. multi-pass approach, combining audio transcription, OCR, and
—Drop first letter if word begins with KN, GN, PN, AE, WR. video feature extraction, achieves high accuracy (99.5%) in dis-
—Convert X to S at start, else to KS. tinguishing between political and non-political content. The subse-
—Transform specific letter combinations. quent analyses of candidate identification and issue-level sentiment

53
 International Journal of Computer Applications (0975 – 8887)
Volume 186 – No.46, November 2024

Table 1. : Performance Metrics for Multi-Pass Classification

Pass Accuracy Precision Recall F1-Score AUC-ROC

Audio Only 0.980 0.975 0.985 0.980 0.995
Audio + OCR 0.992 0.990 0.994 0.992 0.998
Audio + OCR + Video 0.995 0.994 0.996 0.995 0.999

Table 2. : Top 10 Audio TF-IDF Features (4) Insights into the most indicative features for political content
identification
Feature IDF Coefficient
While the system demonstrates significant improvements over ex-
vote 2.809 3.8929
isting methods, there are several avenues for future work:
run 3.016 2.7579
congress 3.115 2.5478 (1) Incorporating more advanced deep learning techniques for
approve 2.943 2.4348 video analysis, such as 3D convolutional neural networks for
message 2.929 2.4309 spatiotemporal feature extraction [29].
state 3.308 2.2994
(2) Expanding the system to handle multi-lingual content, en-
fight 3.107 2.2792
abling analysis of political ads in diverse linguistic contexts.
city 3.533 2.1901
conservative 3.654 2.0535 (3) Developing methods to track the evolution of political issues
border 3.760 1.9333 and sentiment over time, providing insights into changing cam-
paign strategies and public opinion.
(4) Investigating the potential of transfer learning to adapt the
Table 3. : Top 10 OCR TF-IDF Features
model to different political systems and cultural contexts
Feature IDF Coefficient As the landscape of political advertising continues to evolve, par-
paid 2.211 5.4749 ticularly in digital spaces, the need for sophisticated, automated
congress 2.744 5.3974 analysis tools grows increasingly crucial. This work contributes to
vote 3.328 2.9487 this field by providing a comprehensive, accurate, and interpretable
senate 4.142 2.2950 system for political ad analysis, with potential applications in con-
approved 3.323 2.1754 tent moderation, campaign strategy, and political communication
district 3.792 2.0947 research.
elect 4.189 2.0017
state 3.397 1.9988
democrat 3.721 1.7732
10. REFERENCES
vegas 3.825 1.7139 [1] W. Wu, H. Wang, Y. Ye, and Z. Zhang. A comprehensive
survey of video-based action recognition using deep learning:
Dataset, method, and challenge. IEEE Transactions on Neural
Table 4. : Top 8 Key Phrases for Candidate Identification
Networks and Learning Systems, 33(9):4104–4124, 2022.
Key Phrase Pattern Weight Direction [2] eMarketer. Worldwide digital ad spending 2021. eMarketer
Insider Intelligence, 2021.
I approve this message 1.0 Before
[3] Advertising Analytics and Cross Screen Media. 2024 political
I’m [. . . ] and I approve this message 1.0 Before advertising outlook, 2023.
paid for by (. . . ) 0.85 After
[4] S. Dhakal. Deep learning approach for political video adver-
authorized by (. . . ) 0.8 After tisement classification. In Proceedings of the 2019 3rd Inter-
sponsored by (. . . ) 0.8 After national Conference on Deep Learning Technologies (ICDLT
vote for (. . . ) 0.75 After 2019), pages 31–35. Association for Computing Machinery,
2019.
elect (. . . ) 0.75 After
[5] S. B. Belhaouari, A. Alshabani, A. T. Azar, and E. Al-
re-elect (. . . ) 0.75 After
mazrouei. A comprehensive review of deep learning models
for video classification. Applied Sciences, 13(2):890, 2023.
provide valuable insights into the content and tone of political mes- [6] J. W. Grigsby and E. F. Fowler. Political advertising in the
saging. digital age: The political ideology of ads on facebook. Politi-
Key contributions of this work include: cal Communication, 37(6):785–809, 2020.
(1) A novel multi-modal approach to political ad classification [7] Stand by your ad act of 2002, 47 u.s.c. § 315 note (2002),
2002.
(2) Robust methods for candidate verification using fuzzy match- [8] OpenAI. Whisper: Openai’s automatic speech recognition
ing and official databases model [computer software], 2022. [Link]
(3) Granular issue detection and sentiment analysis capabilities openai/whisper.

54
 International Journal of Computer Applications (0975 – 8887)
Volume 186 – No.46, November 2024
H. Jelodar, Y. Wang, C. Yuan, X. Feng, X. Jiang, Y. Li, and
[9] H. Schu¨tze, C. D. Manning, and P. Raghavan. Introduction L. Zhao. Latent dirichlet allocation (lda) and topic modeling:
to information retrieval, volume 39. Cambridge University models, applications, a survey. Multimedia Tools and Appli-
Press, 2008. cations, 78:15169–15211, 2019.
[10] A. Wendland, M. Zenere, and J. Niemann. Introduction to text [26] V. Sanh, L. Debut, J. Chaumond, and T. Wolf. Distilbert, a
classification: impact of stemming and comparing tf-idf and distilled version of bert: smaller, faster, cheaper and lighter,
count vectorization as feature extraction technique. In Sys- 2019. [Link]
tems, Software and Services Process Improvement: 28th Eu-
[27] S. Loria. textblob documentation, 2018. https://
ropean Conference, EuroSPI 2021, Krems, Austria, Septem-
[Link]/en/dev/.
ber 1–3, 2021, Proceedings 28, pages 289–300. Springer In-
ternational Publishing, 2021. [28] D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri.
Learning spatiotemporal features with 3d convolutional net-
[11] JaidedAI. Easyocr: Ready-to-use optical character recog-
works. In Proceedings of the IEEE international conference
nition with 80+ supported languages [computer software],
on computer vision, pages 4489–4497, 2015.
2020. [Link]
[12] M. K. Asha Paul, J. Kavitha, and P. A. Jansi Rani. Key-frame
extraction techniques: A review. Recent Patents on Computer
Science, 11(1):3–16, 2018.
[13] Joe Biden. Joe biden for president - joe biden: Keep up the
fight. YouTube video, 2020. Accessed: 2024-09-10.
[14] G. Bradski. The opencv library [computer software], 2000.
[Link]
[15] L. Cuimei, Q. Zhiliang, J. Nan, and W. Jianhua. Human face
detection algorithm via haar cascade classifier combined with
three additional classifiers. In 2017 13th IEEE interna- tional
conference on electronic measurement & instruments
(ICEMI), pages 483–487. IEEE, 2017.
[16] OpenCV Team. Face detection using haar cascades.
[Link]
js_face_detection.html, 2023. Accessed: [Insert access
date here].
[17] T. Fushiki. Estimation of prediction error by using k-
fold cross-validation. Statistics and Computing, 21:137–146,
2011.
[18] Y. Li, R. Krishnamurthy, S. Raghavan, S. Vaithyanathan, and
H. V. Jagadish. Regular expression learning for information
extraction. In Proceedings of the 2008 conference on empir-
ical methods in natural language processing, pages 21–30,
2008.
[19] B. Mohit. Named entity recognition. In Natural language
processing of semitic languages, pages 221–245. Springer
Berlin Heidelberg, 2014.
[20] M. Cayrol, H. Farreny, and H. Prade. Fuzzy pattern matching.
Kybernetes, 11(2):103–116, 1982.
[21] L. Yujian and L. Bo. A normalized levenshtein distance met-
ric. IEEE Transactions on Pattern Analysis and Machine In-
telligence, 29(6):1091–1095, 2007.
[22] Y. Wang, J. Qin, and W. Wang. Efficient approximate entity
matching using jaro-winkler distance. In International con-
ference on web information systems engineering, pages 231–
239. Springer International Publishing, 2017.
[23] C. Snae. A comparison and analysis of name matching algo-
rithms. International Journal of Computer and Information
Engineering, 1(1):107–112, 2007.
[24] Federal Election Commission. Candidate and commit- tee
viewer [data set]. [Link]
candidates/.
[25] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss,
V. Dubourg, J. Vanderplas, et al. Scikit-learn: Machine
learning in python. Journal of Machine Learning Research,
12:2825–2830, 2011.

55
IJCATM : [Link]