A Novel LSTM Based Model for Sentiment Detection in

Hindi-English Code-Switched Texts

Namitha Shambhu Bhat

, Kuldeep Sambrekar

and Shridhar Allagi

Computer Science and Engineering, KLS Gogte Institute of Technology, Belagavi, India

Computer Science and Engineering, KLE Institute of Technology, Hubli, India

Keywords: Sentiment Analysis, Code-Switching, Language Identification, LSTM, Switch Point Detection.

Abstract: Sentiment analysis in multilingual conversations is laborious yet important in understanding emotions and

opinions expressed across multiple languages and cultures. Code-switching, a prevalent technique poses

challenges due to linguistic diversity, cultural nuances, and contextual dependencies. The research in this

article provides an LSTM-based framework for sentiment analysis in Hindi-English code-switched text,

addressing the challenges of multilingual content in social media. The methodology adopted in this research

incorporates three key components: language-specific encoders to obtain linguistic patterns, a switcher

module for understanding language transitions, and a sentiment analysis module for extracting sentiment

within a multilingual text. A Hindi-English dataset containing 4,954 samples with positive, neutral, and

negative sentiments is used for training and evaluation. The model achieves an overall accuracy of 89.9%,

and an F1-score of 0.9 across all sentiments investigated. This work contributes substantially to multilingual

sentiment analysis, eliminating the shortcomings of conventional approaches and offering a robust method

for analyzing complex code-switched text.

1 INTRODUCTION

Multilingual Code-Switching (MCS) is a prevalent

phenomenon in global societies, where individuals

switch between two or more languages within a single

conversation(Myers-Scotton 1993). This linguistic

phenomenon is widespread in multicultural

communities, social media, and online forums (Plaza-

del-Arco et al. 2021) (AlGhamdi et al. 2016).

Sentiment analysis (SA) in MCS texts is crucial for

understanding public opinions, emotions, and

intentions (Pang and Lee 2008). However, MCS

poses significant challenges for SA due to its complex

linguistic patterns, cultural nuances, and contextual

dependencies. The increasing volume of MCS text

data necessitates effective SA tools to extract valuable

insights. Traditional SA approaches focus on

monolingual text analysis, neglecting the

complexities of MCS (Jamatia et al. 2020) (Zhu et al.

2022) (Tan, Lee, and Lim 2023). Recent studies have

https://orcid.org/0009-0007-8945-4197

https://orcid.org/0000-0002-3542-1716

https://orcid.org/0000-0001-7502-0741

addressed MCS SA using rule-based and machine-

learning approaches (Kim n.d.; P and Mahender

2024), (Ullah et al., 2022). However, these methods

have limitations, such as relying on hand-crafted

features or requiring large annotated datasets.

Deep learning techniques have shown promise in

SA (Dutta, Agrawal, and Kumar Roy 2021) (Santos

and Gatti 2014). However, their application to MCS

SA is still developing. This research aims to bridge

this gap by proposing a novel deep-learning approach

for MCS SA. The complexity of MCS texts arises

from various factors, including Language

identification (S. D. Das et al., 2019)(D. Das &

Petrov, 2011), contextual understanding (Code-

switching n.d.) (Gardner-Chloros 2009), and cultural

influences (Albahoth et al., 2024; Hofstede, 2001).

Current approaches to MCS SA mainly focus on

Rule-based methods that use hand-crafted rules and

linguistic features (Agüero-Torales et al., 2021), and

Machine learning leveraging supervised learning

techniques with annotated datasets (Chakravarthi et

422

Bhat, N. S., Sambrekar, K. and Allagi, S.

A Novel LSTM-Based Model for Sentiment Detection in Hindi- English Code-Switched Texts.

DOI: 10.5220/0013593600004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 422-428

ISBN: 978-989-758-763-4

al., 2020). However, these approaches have

limitations wherein Hand-crafted features may not

capture complex linguistic patterns., Annotated

datasets are scarce, especially for low-resource

languages, Cultural and contextual nuances are often

overlooked.

This research is further organised as follows: We

have presented the related prior research in section 2,

and the methodology for assimilating the gaps and

required solutions is formulated and presented in

section 3. Section 4 showcases the experimental

results obtained using the datasets and the processing

techniques mentioned in the method with elaborate

discussions. In section 5, we provide compelling

concluding remarks highlighting the challenges of

sentiment analysis and outline significant

opportunities for future research and improvement.

2 RELATED STUDIES

Research on computational models for code-mixing

is limited due to the scarcity of this phenomenon in

conventional text corpora, making data-greedy

approaches difficult to apply. Recent studies have

highlighted the complexity and importance of

sentiment analysis in multilingual and code-switched

contexts. For instance, Sharma et al. (Sharma et al.,

2023a) developed a sentiment analysis system for

code-switched data using a late fusion approach

combining two transformers, which showed

promising results on English-Hindi and English-

Spanish datasets (Sharma et al., 2023b). Similarly,

Vilares et al. (Vilares et al., 2016) explored sentiment

analysis on monolingual, multilingual, and code-

switching Twitter corpora, emphasizing the unique

challenges posed by code-switching. Additionally,

Mamta and Ekbal (Mamta and Ekbal 2025) proposed

a transformer-based multilingual joint learning

framework for code-mixed and English sentiment

analysis, demonstrating improved performance

through shared-private, multi-task learning.

Furthermore, a study by Yekrangi and Abdolvand

(Yekrangi and Abdolvand 2021) they were focused

on augmenting sentiment prediction capabilities for

code-mixed languages, including English and Roman

Urdu, using robust transformer-based algorithms.

These studies underscore the growing recognition of

code-switching in sentiment analysis and the need for

advanced models to handle the linguistic diversity

present in social media content. Another study by

Kumar et al. (Kumari and Kumar 2021) revealed the

impact of code-switching on sentiment analysis,

highlighting the need for specialized models to handle

the linguistic diversity in social media content.

Additionally, a recent study by Patwa et al. (Patwa et

al., 2020) examined the effectiveness of various

machine learning models for sentiment analysis in

code-mixed social media text, finding that

transformer-based models outperformed traditional

approaches. An exhaustive study of the prior research

is presented in Table 1.

These studies underscore the growing recognition

of code-switching in sentiment analysis and the need

for advanced models to handle the linguistic diversity

present in social media content. In the research

presented herein, we have used multiple techniques to

address the issues related to sentiment analysis and

achieve an optimum result that bridges the gap

mentioned in the related research. This research

proposes a novel deep learning approach for MCS

SA, incorporating a combination of three techniques

namely,

 Language-specific encoders to capture

linguistic patterns.

 A switcher module to detect language

switching.

 A sentiment analysis module to capture

contextual nuances.

This approach effectively utilizes deep learning

techniques, aiming to improve accuracy in MCS SA,

reduce reliance on hand-crafted features, and enable

more effective sentiment analysis in multilingual

settings.

3 METHODOLOGY

This research proposes a novel deep-learning

approach for sentiment analysis in multilingual code-

switching text shown in Figure 1.

Figure 1: Proposed Methodology for Multilingual

Sentiment Analysis.

In the dataset preparation and preprocessing a

Hindi-English mixed dataset samples are collected

from social media. Sentiment labels such as Positive,

Neutral, and Negative are annotated to the dataset. A

Input Text

Preprocessing

(Tokenization,

Stopword Removal)

Language Identification

(LangID, LSTM)

Switcher Module

(Determines Language

Switch)

Sentiment Analysis

Module

(Concatenates encoder

outputs)

Sentiment

Classification

A Novel LSTM-Based Model for Sentiment Detection in Hindi- English Code-Switched Texts

423

total of 4954 samples are used, consisting of a mix of

1850 Positive, 1700 Neutral, and 1404 Negative

samples. Tokenization was executed to split the input

text into individual words for further analysis.

Table 1: Previous Research on Multilingual Sentiment Analysis.

Study Methodology Language(s) Dataset Accuracy/F-

score

Limitations Year

(Kasmuri et

al. 2020)

Rule Based Malay-

English

Social media 0.87 (F1-score) Limited to specific

language pairs,

relies on hand-

crafted rules

2020

(Younas et

al. 2020)

State-of-the-art

Deep learning

models

(mBERT, XLM-

Roman

Urdu-

English

Social media mBERT:

69%

(Accuracy)

0.69 (F1-score)

XLM-R:

71%

(Accuracy)

0.71

(

F1-score

)

Lack of exploration

into the contextual

nuances and cultural

references in the

code-mixed text that

could influence

sentiment

interpretation.

2020

(Sharma,

Chinmay,

and Sharma

2023b)

mBERT-BERT

Late fusion

models

English-

Hindi

Social media 0.6124 (F1-

score)

Focuses on widely

spoken languages.

Low resource

languages need to

e ex

lore

2023

(Adel et al.

2013)

RNN Language

Model

English-

German

Speech Data - Focuses on

language modeling,

may not generalize

well to sentiment

analysis

2013

(Santos and

Gatti 2014)

CNN Sentiment

Analysis

English-

Portuguese

Short Texts 0.925 (F1-

score)

May not perform

well with longer

texts or different

lan

airs

2014

(Klementiev,

Titov, and

Bhattarai

2012)

Cross-lingual

Sentiment

English-

Spanish-

Multilingual 0.859

(Accuracy)

Assumes parallel

corpora availability,

may not handle

code-switching

within sentences

2012

(Glorot,

Bordes, and

Bengio

2011)

Domain

Adaptation

English-

French

Multilingual 0.885

(Accuracy)

Requires target

domain labeled

data, may not adapt

well to unseen

domains

2011

(You, Jin,

and Luo

2017)

Visual

Sentiment

Analysis

English Image Data 0.921

(Accuracy)

Limited to visual

features, may not

capture textual

sentiment

2016

(Zadeh et al.

2017)

Tensor Fusion

Network for

Multimodal

Sentiment

English Multimodal 0.942 (F1-

score)

Requires

multimodal data,

may face challenges

in integrating

multiple modalities

2017

Stopwords like “is”, “and”,” or”, “the”, “aur”, “mein”

etc. which did not carry any significant meanings

were removed from the tokenized words. This step

reduces the dimensionality whilst retaining the

quality of the data. Special characters such as emojis,

hashtags and punctuations are omitted to prevent

trivial features from affecting the performance of the

model.

LangID, a prebuilt library is used to tag each token

with its language English or Hindi. All English words

are converted to lowercase and the Hindi words to

Unicode for a uniformity in formats. The dataset is

INCOFT 2025 - International Conference on Futuristic Technology

424

later split into training and testing sets at an 80:20

ratio. The Architecture of the model was built using:

3.1 Language Identification Module

The Pre-trained LangID embeddings are used to

detect language-specific patterns and switching

between Hindi and English. The Language-Specific

Encoders contain separate input layers which are

defined for Hindi and English and embedding layers

are created to convert integer-encoded words into

dense vectors. The dimension of embedding vectors

in this layer is set to 100 to convert input sequences

into embedded representations. A single layer of

LSTM with 64 units is implemented to capture

sequential dependencies. This step helps in

understanding where and why a language switch

occurs. A dropout layer with a dropout rate of 0.2 is

created for regularization. It helps prevent overfitting

by randomly dropping 20% of the units during

training.

3.2 Switcher Module

This module is applied to identify where the language

changes from Hindi to English or vice versa within a

sentence. The execution is carried out in two steps

namely Sequence Labelling with LSTM-based

Network and Training with Annotated Switch Points.

The former step involves assigning a label to each

token in the sentence and the latter step is used to train

the model using a labelled dataset where each token

was annotated with its respective language. For the

switcher module, labels denote the language of the

token i.e. English or Hindi.

3.3 Sentiment Analysis Module

The output from the language-specific encoders is fed

to the sentiment analysis module to predict the

sentiment of the input text. This module integrates

information from both the Hindi and English parts of

the sentence. The outputs from the Hindi and English

language-specific encoders are concatenated. These

outputs are high-dimensional vectors (embeddings)

representing the semantic meaning of the text in both

languages. A fully connected dense layer with 64

hidden units is applied to the concatenated vector.

Rectified Linear Unit (ReLU) introduces non-

linearity, allowing the model to learn complex

patterns. A final Dense Layer with Softmax

Activation predicts the sentiment class:

Positive (e.g., "I love this movie!"),

Neutral (e.g., "The day was okay."),

Negative (e.g., "I hated the food.").

Softmax activation converts raw scores into

probabilities, ensuring they sum to 1. The class with

the highest probability is chosen as the sentiment

prediction.

3.4 Model Compilation and Training

Adam optimizer with a learning rate of 0.001 is used

to compile the model. Sparse Categorical Cross

entropy is applied as the loss function to understand

the difference between the true class label (as an

integer) and the predicted probability distribution

over the classes output by the model. The model is

trained using 3964 samples and the remaining 990

samples are used for validation with batch size 32 for

20 epochs.

3.5 Evaluation Metrics

Evaluation Metrics used to measure the model’s

performance are accuracy, precision, recall and F1-

score. Accuracy measures the percentage of

appropriately classified samples out of the total

samples.

𝐴

𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =





(1)

Precision measures the proportion of correctly

predicted positive samples out of all samples

predicted as positive.

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =





(2)

Recall measures the proportion of actual positive

samples that were correctly predicted as positive.

𝑅𝑒𝑐𝑎𝑙𝑙 =





(3)

F1-score is the harmonic mean of Precision and

Recall, providing a balance between the two.

𝐹1 = 2×

  









(4)

The results of these metrics are discussed and

presented in the next section.

A Novel LSTM-Based Model for Sentiment Detection in Hindi- English Code-Switched Texts

425

4 EXPERIMENTAL RESULTS

The results of the proposed approach to the sentiment

analysis in the English-Hindi code-switching dataset

are shown using the confusion matrix in Table 2. The

positive and negative classes attain marginally higher

performance metrics, implying the model's confidence

in determining these sentiments. The neutral class,

exhibits slightly lower recall, leading to incidental

misclassifications into positive or negative categories.

Table 2: Confusion Matrix of the Proposed Approach.

Class Predicted

Positive

Predicted

Neutral

Predicted

ative

True

Positive

1700 100 50

True

Neutral

120 1500 80

True

ative

70 76 1258

It is observed that the model has achieved strong

bilingual performance with an overall accuracy of

89.9% with better recall, precision and an average F1

score of 0.9 calculated using equations 1 to 4

respectively for positive, neutral, and negative classes

shown in Table 3.

Table 3: Precision, Recall, And F1-Score for Each Class.

Class Precision Recall F1-Score

Positive 0.899 0.919 0.909

Neutral 0.895 0.882 0.889

Negative 0.906 0.896 0.901

Overall Model

Performance

0.9 0.899 0.899

The ROC curve represented in Figure 2 reveals

that both the positive and negative classes possess

high AUC values of 0.93, indicating outstanding

separability. The neutral class showcases a robust

ability to distinguish its predictions with an

inconsiderable lower AUC of 0.91. The curves rise

steeply, implying that the model achieves high True

Positive Rates (TPR) with minimal False Positive

Rates (FPR). All curves observed in Figure 2 overlie

the diagonal chance line, confirming the model's

effectiveness over random assumptions. In principle,

the model performs reliably, with balanced

classification capabilities across all sentiment

categories.

5 CONCLUSION AND FUTURE

IMPLICATIONS

This research demonstrates a notable advancement in

sentiment analysis for Hindi-English code-switched

texts through a novel LSTM-based model that excels

in both accuracy and F1-score, key metrics for

evaluating classification models. Achieving an

overall accuracy of 89.9%, the model correctly

classifies a high proportion of sentiments across

positive, neutral, and negative categories, showcasing

its reliability in real-world applications. The F1-score,

a harmonic mean of precision and recall, underscores

the model's balanced performance by ensuring both

correctness and completeness in predictions. With an

F1-score of 0.9, the study highlights the model's

proficiency in capturing sentiments even in

linguistically complex and contextually varied code-

switched text, reflecting its ability to minimize false

positives and negatives effectively. This balance is

particularly important in multilingual sentiment

analysis, where nuances in language transitions can

pose significant challenges. By achieving such

metrics, the study not only validates the robustness of

its methodology but also establishes a strong

foundation for addressing broader applications,

including low-resource language pairs and more

complex multilingual contexts, while encouraging

further refinements in accuracy and sentiment

detection capabilities.

Further research can be extended by exploring

additional low-resource language pairs, improving

neutral sentiment detection, and involving advanced

deep learning architectures like transformers to

enhance performance and adaptability to distinct

multilingual contexts.

Figure 2: ROC-AUC Curve.

INCOFT 2025 - International Conference on Futuristic Technology

426

REFERENCES

Adel, H., Vu, N. T., Kraus, F., Schlippe, T., Li, H., &

Schultz, T. (2013). Recurrent neural network language

modeling for code switching conversational speech.

ICASSP, IEEE International Conference on Acoustics,

Speech and Signal Processing - Proceedings, 8411–

8415. https://doi.org/10.1109/ICASSP.2013.6639306

Agüero-Torales, M. M., Abreu Salas, J. I., & López-

Herrera, A. G. (2021). Deep learning and multilingual

sentiment analysis on social media data: An overview.

Applied Soft Computing, 107, 107373.

https://doi.org/10.1016/J.ASOC.2021.107373

Albahoth, Z. M., Jabar, M. A. bin A., & Jalis, F. M. B. M.

(2024). A Systematic Review of the Literature on Code-

Switching and a Discussion on Future Directions.

International Journal of Academic Research in Business

and Social Sciences, 14(2).

https://doi.org/10.6007/ijarbss/v14-i2/20452

AlGhamdi, F., Molina, G., Diab, M., Solorio, T., Hawwari,

A., Soto, V., & Hirschberg, J. (2016). Part of Speech

Tagging for Code Switched Data. EMNLP 2016 - 2nd

Workshop on Computational Approaches to Code

Switching, CS 2016 - Proceedings of the Workshop,

98–107. https://doi.org/10.18653/V1/W16-5812

Chakravarthi, B. R., Jose, N., Suryawanshi, S., Sherly, E.,

& Mccrae, J. P. (2020). A Sentiment Analysis Dataset

for Code-Mixed Malayalam-English.

https://pypi.org/project/langdetect/

Code-switching. (n.d.). Retrieved November 11, 2024, from

https://www.cambridge.org/core/books/codeswitching/

11E359BFC45F331519170EE425117736

Das, D., & Petrov, S. (2011). Unsupervised part-of-speech

tagging with bilingual graph-based projections. ACL-

HLT 2011 - Proceedings of the 49th Annual Meeting of

the Association for Computational Linguistics: Human

Language Technologies, 1, 600–609.

Das, S. D., Mandal, S., & Das, D. (2019). Language

identification of Bengali-English code-mixed data

using character & phonetic based LSTM models. ACM

International Conference Proceeding Series, 60–64.

https://doi.org/10.1145/3368567.3368578

Dutta, S., Agrawal, H., & Kumar Roy, P. (2021). Sentiment

Analysis on Multilingual Code-Mixed Kannada

Language. https://dravidian-

codemix.github.io/2021/index.html

Gardner-Chloros, P. (2009). Code-switching. Code-

Switching, 1–242.

https://doi.org/10.1017/CBO9780511609787

Glorot, X., Bordes, A., & Bengio, Y. (2011). Domain

adaptation for large-scale sentiment classification: A

deep learning approach. Proceedings of the 28th

International Conference on Machine Learning, ICML

2011, 1, 513–520.

Hofstede, G. (2001). Culture’s Consequences: Comparing

Values, Behaviors, Institutions, and Organizations

Across Nations. Culture’s Consequences: Comparing

Values, Behaviors, Institutions, and Organizations

Across Nations, 41(7), 861–862.

https://doi.org/https://doi.org/10.1016/S0005-

7967(02)00184-5

Jamatia, A., Swamy, S. D., Gambäck, B., Das, A., &

Debbarma, S. (2020). Deep Learning Based Sentiment

Analysis in a Code-Mixed English-Hindi and English-

Bengali Social Media Corpus. International Journal on

Artificial Intelligence Tools, 29(5).

https://doi.org/10.1142/S0218213020500141

Kasmuri, E., Fakulti, H. B., Maklumat, T., Komunikasi, D.,

& Maklumat, F. T. (2020). Segregation of Code-

Switching Sentences using Rule-Based Technique. Int.

J. Advance Soft Compu. Appl, 12(1).

Kim, H. (n.d.). 1 Semi Annual Edition.

Klementiev, A., Titov, I., & Bhattarai, B. (2012). Inducing

crosslingual distributed representations of words. 24th

International Conference on Computational Linguistics

- Proceedings of COLING 2012: Technical Papers,

December, 1459–1474.

Kumari, J., & Kumar, A. (2021). Offensive Language

Identification on Multilingual Code Mixing Text.

CEUR Workshop Proceedings, 3159, 643–650.

https://github.com/Abhinavkmr/Dravidian-hate-

speech.git

Mamta, & Ekbal, A. (2025). Quality achhi hai (is good),

satisfied! Towards aspect based sentiment analysis in

code-mixed language. Computer Speech and Language,

89. https://doi.org/10.1016/j.csl.2024.101668

Myers-Scotton, C. (1993). Common and Uncommon

Ground: Social and Structural Factors in

Codeswitching. Language in Society, 22(4), 475–503.

https://doi.org/10.1017/S0047404500017449

P, K. V., & Mahender, C. N. (2024). A Named Entity

Recognition System for the Marathi Language.

JOURNAL OF ADVANCED APPLIED SCIENTIFIC

RESEARCH, 6(3).

https://doi.org/10.46947/JOAASR632024937

Pang, B., & Lee, L. (2008). Opinion mining and sentiment

analysis. Found Trends Inf Retr, 2(1–2), 1–135.

https://doi.org/10.1561/1500000011

Patwa, P., Aguilar, G., Kar, S., Pandey, S., Srinivas, P. Y.

K. L., Gambäck, B., Chakraborty, T., Solorio, T., &

Das, A. (2020). SemEval-2020 Task 9: Overview of

Sentiment Analysis of Code-Mixed Tweets. 14th

International Workshops on Semantic Evaluation,

SemEval 2020 - Co-Located 28th International

Conference on Computational Linguistics, COLING

2020, Proceedings, 774–790.

https://doi.org/10.18653/v1/2020.semeval-1.100

Plaza-del-Arco, F. M., Molina-González, M. D., Ureña-

López, L. A., & Martín-Valdivia, M. T. (2021). No

Title. 166(114), 120.

https://doi.org/10.1016/j.eswa.2020.114120

Santos, C. dos, & Gatti, M. (2014). Deep Convolutional

Neural Networks for Sentiment Analysis of Short Texts

(pp. 69–78). https://aclanthology.org/C14-1008

Sharma, G., Chinmay, R., & Sharma, R. (2023a). Late

Fusion of Transformers for Sentiment Analysis of

Code-Switched Data. Findings of the Association for

Computational Linguistics: EMNLP 2023, 6485–6490.

A Novel LSTM-Based Model for Sentiment Detection in Hindi- English Code-Switched Texts

427

https://doi.org/10.18653/V1/2023.FINDINGS-

EMNLP.430

Sharma, G., Chinmay, R., & Sharma, R. (2023b). Late

Fusion of Transformers for Sentiment Analysis of

Code-Switched Data. Findings of the Association for

Computational Linguistics: EMNLP 2023, 6485–6490.

https://doi.org/10.18653/v1/2023.findings-emnlp.430

Ullah, A., Khan, S. N., & Nawi, N. M. (2022). Review on

sentiment analysis for text classification techniques

from 2010 to 2021. Multimedia Tools and Applications

2022 82:6, 82(6), 8137–8193.

https://doi.org/10.1007/S11042-022-14112-3

Vilares, D., Alonso, M. A., & Gómez-Rodríguez, C. (2016).

En-ES-CS: An English-Spanish code-switching twitter

corpus for multilingual sentiment analysis. Proceedings

of the 10th International Conference on Language

Resources and Evaluation, LREC 2016, 4149–4153.

http://grupolys.org/

Yekrangi, M., & Abdolvand, N. (2021). Financial markets

sentiment analysis: developing a specialized Lexicon.

Journal of Intelligent Information Systems, 57(1), 127–

146. https://doi.org/10.1007/S10844-020-00630-

9/TABLES/2

You, Q., Jin, H., & Luo, J. (2017). Visual Sentiment

Analysis by Attending on Local Image Regions.

Proceedings of the AAAI Conference on Artificial

Intelligence, 31(1), 123–125.

https://doi.org/10.1609/aaai.v31i1.10501

Younas, A., Nasim, R., Ali, S., Wang, G., & Qi, F. (2020).

Sentiment analysis of code-mixed roman Urdu-English

social media text using deep learning approaches.

Proceedings - 2020 IEEE 23rd International Conference

on Computational Science and Engineering, CSE 2020,

66–71. https://doi.org/10.1109/CSE50738.2020.00017

Zadeh, A., Chen, M., Cambria, E., Poria, S., & Morency, L.

P. (2017). Tensor Fusion Network for Multimodal

Sentiment Analysis. EMNLP 2017 - Conference on

Empirical Methods in Natural Language Processing,

Proceedings, 1103–1114.

https://doi.org/10.18653/v1/d17-1115

INCOFT 2025 - International Conference on Futuristic Technology

428