A Comparative Study of BERT, RoBERTa, and Bi-LSTM for Sentiment Analysis of Social Media Data in Diverse Domains

A Comparative Study of BERT, RoBERTa, and Bi-LSTM for Sentiment Analysis of Social Media Data in Diverse Domains

Sunny Suresh Nahar, Swapna Kulkarni, Dr. Prashant Prakashrao Agnihotri

School of Technology, Swami Ramanand Teerth Marathwada University, Sub-Campus, Latur, Maharashtra, India

Abstract – Sentiment Analysis (SA) has become a very critical activity for gaining valuable information from the huge amount of user-generated content that is available on social media sites. Yet, sentiment classification has proved to be difficult due to reasons like language informality, contextual uncertainty, noise in the text, abbreviations, emojis, etc. The goal of this research is to conduct a comparative assessment of Bi-directional Encoder Representations Transformer (BERT) and Robustly Optimized BERT Pretraining Approach (RoBERTa) models for SA on various social media domains through one common experimental setup. The suggested experimental setup would include text processing, feature extraction using Word2Vec, feature selection using Mutual Information (MI), and training and testing of the models on the Sentiment140 Dataset with 1.6 million tweets. Results from experiments indicate that the RoBERTa-based framework attains the highest accuracy of 98.56% along with the highest precision, recall, F1 score, and 5.46 seconds of training time and outperforms Bi-LSTM and BERT in terms of classification efficacy and computational efficiency. The novelty of the proposed study lies in ensuring a fair comparison through identical experimental settings while integrating Word2Vec- based feature representation with Mutual Information-based feature selection to enhance classification performance. The suggested framework is a highly reliable and scalable solution for performing accurate sentiment classification and helps in selecting the right model.

Keywords BERT; Bi-LSTM; Mutual Information; RoBERTa; Sentiment Analysis

1. INTRODUCTION

   The rapid surge of social media utilization has yielded an unprecedented volume of user-generated content, conveying public opinions within multiple sectors, including medicine, politics, finance, marketing, and entertainment [1], [2]. Consequently, SA has become an important task that can help to extract useful information from such content for an organization to track public opinion or to forecast trends [3- 5]. Nonetheless, social media text SA has great difficulties. The use of informal and noisy content such as slang, abbreviations, emojis, and poor grammar is commonly seen in social media content [6], [7]. The performance of the model gets affected due to the high variability in domains. Sarcasm, contextual ambiguity, multilingual or code-mixed inputs, class imbalance, etc., can severely affect classification robustness [8], [9]. Models that can capture deep, appropriate semantics and long -range dependencies are needed to tackle this complexity [10], [11]. Fig. 1 illustrates the basic steps of SA.

   U U U U

   Text

Social Media

Data

Classification

Data

Fig. 1. Basic steps of SA [12]

Traditional Machine Learning (ML) algorithms [13], [14] that adopt Bag-of-Words and TF-IDF rely heavily on hand- crafted features, despite being more efficient than other types of models [15-18]. Nevertheless, they cannot integrate contextual relations, word dependency owing to order, and circa semantics; hence, they do not generalize well across domains of different types [19-21]. Deep Learning’s (DL) recent advancements have substantially enhanced the performance of sentiment categorization tasks, as an expansion of Bi-LSTM networks [22], which may learn the dependencies in both directions at the same time [23], [24]. Transformers’ architectural design has introduced new possibilities for natural language processing through its dual mechanisms of self-attention and pretraining, which enable better contextual comprehension and domain adaptation abilities for users [25], [26]. The different architectural design elements, which include their capacity to represent context, their total number of parameters, and their processing needs, can determine how well they perform with various data sets [27]. The existing research requires domain experts to understand different model selection methods because professional fields require accuracy and generalization, and computational costs to be balanced [28]. The current study carries out an extensive evaluation that compares the performance of the three DL models for social media sentiment data analysis in different domains. The study gives a comprehensive assessment of performance metrics, generalization ability, and computational speed on various social media platforms, which is different from other effort that employs one architecture on domain-specific datasets. This research proposes an evaluation methodology that directly compares Bi-LSTM models under a standardized test system, along with the transformer-based models, BERT and RoBERTa, on multiple dimensions. Comparison between recurrent networks and self-attention networks are possible from identical setups and evaluations, which reveal how each one learns context. This research aims to compare the BERT and RoBERTa models that analyze the

sentiment of social media data across different industries. This research includes these essential contributions:

An extensive empirical study of Bi-LSTM, BERT, and RoBERTa models for multi-domain sentiment classification. Establishing a common testing platform to measure the robustness and generalization capabilities of the models on various social media platforms.

Testing the trade-offs between the models’ performance and efficiency, including accuracy, training time, and complexity.

Offering valuable recommendations for model selection in real-world SA tasks.

This paper is organized as follows. Section 2 gives an overview of the related literature. The data and techniques used, and the steps followed, are covered in Section 3. Section 4 presented the results and discussion of the research. Lastly, Section 5 shows the conclusion and future directions for research.

1. LITERATURE REVIEW

   In this section, some research works on SA using social media data in different fields are discussed. In 2026, Arularasan et al. [29] presented a new Cycle Generative Adversarial Network (GAN) based Term Memory integrated with SA (CyG-TMSA) approach to improve the performance of opinion mining tasks on social media datasets. The results revealed that the proposed approach achieved 98% accuracy with a processing time of 10.12 s on Amazon and Twitter datasets. Sofi et al. (2026) [30] presented an adaptive ensemble ML classifier for Twitter-based SA by combining four classifiers using voting, bagging, Extreme Gradient Boosting (XGBoost), and stacking with a meta-learner, which performed better than the standalone classifiers in terms of accuracy and robustness. Rahman et al. (2025) [31] combined the hybrid RoBERTa-BiLSTM model to capture remote dependencies. The results of the experiment proved to be successful, as the accuracy for Twitter US Airline was 80.74%, 92.36% for IMDB, and 82.25% for Sentiment140. Siddiqui et al. (2025) [32] conducted a study that comparatively assessed ML and hybrid Bi-LSTM with Word2Vec approach on the Sentiment140 dataset, and it found that the hybrid Bi-LSTM architecture outclassed traditional ML models in terms of overall classification accuracy based on confusion matrix evaluation. Deng and Liu (2025) [33] proposed a RoBERTaBiLSTMAttention Model incorporated with supplementary sentiment dictionary based on SO-PMI. Their model achieved 91.56% accuracy, 90.87% precision, 91.07% recall, and 91.17% F1- score, suggesting it classification and generalization capability was improved. Karim et al. in 2025 [34] applied support vector machines, Naïve Bayes, and LSTMs in addition to BERT on the Sentiment140 dataset for SA. The results showed that BERT achieved the best accuracy of

   83.48 for predicting trends in consumer behavior. Employing the IMDb and Sentiment140 datasets,

   Elmitwalli and Mehegan (2024) [35] conducted a comprehensive evaluation of several SA models, including lexicon-based, ML, DL, and artificial intelligence. They

   found that the highest F1-scores were achieved by BERT (0.9380 on IMDB; 0.8114 on Sentiment140), while GPT-3 provided the best overall generalization across those partially unannotated COP9 tweets (0.8812). The Dual Joint Classifier (DJC) -based model that Ram (2024) [36] suggested for balancing classes in Twitter SA is a combination of RoBERTa and BERT architectures, with layers of Bidirectional-Gated Recurrent Unit (Bi-GRU) and Bi- LSTM. The DJC model achieved 87.22% accuracy with the US Airline Twitter data set and 93.87% accuracy with the Apple Twitter data set. Tang (2023) [37] created a Transformer-based deep neural network model to perform sentiment classification on the 1.6 million tweets provided by the Sentiment140 dataset. Experimental results demonstrated significant accuracy and robustness improvements, indicating that Transformer architectures have a better capability to handle contextual dependencies with large-scale social media data. The Bi-LSTM-based methodology given by Arbane et al. (2023) [38] was utilized to classify public sentiment regarding COVID-19 technology. Additionally, Tan et al. (2022) [39] proposed a hybrid RoBERTa-LSTM architecture that integrated the transformational model’s contextually embedded information with sequentially created LSTM architectures and achieved F1 results of 93% for IMDb, 91% for Twitter US Airline, and 90% for Sentiment140.

   However, some research gaps have been identified despite breakthroughs in the field of SA. Most hybrid and ensemble models were evaluated on standard datasets but were not thoroughly analyzed for cross-domain capabilities and/or evaluated for real-world use [29], [34], [39]. Although BERT and GPT-3 models were noted for better handling of tasks, the complexity and the need for Real-time streaming dealt with the models poor capability [35], [36]. Also, some of the main problems have not been analyzed in-depth, like sarcastic speech detection, multilingual sentiment, low- resource language sentiment, and refined emotional sentiment [32], [38]. Lastly, explainable and interpretable AI was not analyzed at all. This is one of the most important factors for Trust in AI (TnAI) and is significant for improving Trust and Transparency in the sentiment decision support systems.

2. RESEARCH METHODOLOGY

   This section shows the proposed method for multimodal SA, which uses text data for precise classification results, as shown in Fig. 2. The collected data is preprocessed using traditional text processing techniques such as tokenization, removal of stop-words, lemmatization, and normalization. As a feature extraction method, Word2Vec is used with MI for feature selection, allowing identification of the most important features. Bi-LSTM, BERT, and RoBERTa are trained earlier after splitting the data into training and testing data. The training process uses the cross-entropy loss function and Adam’s optimizer. proposed model performance is evaluated using conventional evaluation metrics and is monitored for its steadiness and elasticity for the entire duration of the monitoring process.

   TABLE I: Dataset Summary

   Tokenization

   Removal

   of stop

   Sentiment140 dataset

   Lemmatization and stemming

   Feature Extraction using Word 2Vec

   Normalization

   Training

   Testing

   Feature Selection using MI

   Dataset splitting

   Model Designing and Training

   BI-LSTM

   BERT

   Model evaluation

   Continuous Monitoring

SA

Fig. 2. Design Architecture

1. Dataset Explanation

   The dataset for SA in this study comes from the open- source Kaggle [40] known as Sentiment140, which is specifically meant for the task of SA on Twitter data at scale. The dataset contains roughly 1.6 million tweets, all of which

   Total Records Target

   Tweet_ID User

   TABLE I: DATASET SUMMARY

   Feature Name Description Example

   Number of tweets Sentiment polarity (0 = Negative, 4 = Positive) ID for tweet

   Username

   1,600,000

   0 / 4

   1467858369

   username123

   have been automatically labelled in terms of polarity in

   sentiments and thus are suitable for binary classification

   Tweet_Text

   Raw tweet content

   I love this movie :).

   tasks. Each data entry contains features such as tweet ID, date, username, query feature, and tweet text. This dataset captures real-life usage of social media in terms of language, emoticons, abbreviations, and other noisy features, thus making it a great testbed for evaluating the performance of DL and transformer-based SA methods. Table I displays the structure of the dataset used for training.

2. Data Preprocessing

   To convert noisy and unstructured social media language into a clear, consistent format appropriate for DL models, data preprocessing is an essential step in SA [41], [42]. To convert the unprocessed multimodal textual information into a clear, organized representation that can be used for feature removal, the preprocessing step is carried out in the suggested technique. Let the raw textual corpus be denoted as ‘D =

   {dl, d2, , dM}, where each document diconsists of a sequence of words di. The first step is the removal of stop words, which is performed to eliminate commonly occurring but semantically insignificant words (e.g., is, the, and and). Let S denote the predefined stop-word set. The filtered token set T,(di) is obtained as:

   T,(di) = {tk T(di) tk S} (1)

   This operation reduced dimensionality and improved discriminative feature learning.

   Next, tokenization, in which each document is segmented into individual lexical units (tokens). Formally, a tokenization

   function T() is applied as:

   In the above formula, c is the size of the context window, wt is the target word and wt+j is the immediately next word in the context. SoftMax is used to derive these conditional probabilities:

   exp(vT v )

   WO Wi

   P(w w ) =

   (9)

   T(di) = {tl, t2, , tK}, tk V (2)

   0 i IV

   exp(vTvW )

   W=l W i

   Where V is the vocabulary space and K N after cleaning punctuation and delimiters.

   Subsequently, lemmatization and stemming were applied to reduce morphological variations. For stemming, a function s()truncated tokens to their root forms:

   k s k

   t(s) = (t ) (3)

   Where vwi and vwO are vector interpretations of input and output words, V is the vocabulary size. After training, each word w is represented as a dense vector:

   ew d (10)

   Where d is the embedding dimension. Consequently, each tweet is transformed into an embedding matrix:

   Where t(s)represents the stemmed version of token t . In

   k k Ei = [ew

   , ew

   , , ew

   ] (11)

   contrast, lemmatization used a vocabulary-based mapping function l()to obtain the canonical lemma:

   k l k

   t(l) = (t POS) (4)

   Where POS denotes the part-of-speech tag, ensuring linguistically valid base forms (e.g., running run). Finally, normalization was conducted to standardize token representations. This included lowercasing, removal of special characters, and numerical standardiztion. Formally, a normalization operator N()was applied:

   t(n) = N(t ) (5)

   il i2 in

   This dense semantic representation captures contextual similarity and reduces dimensional sparsity, thereby enhancing the capability of subsequent Bi-LSTM layers to model sequential dependencies.

   D. Feature Selection using MI

   Feature selection is carried out to determine the most informative features that play a significant role in sentiment classification with the aim of eliminating redundancy and complexity [45]. In this research, MI is used to determine the

   k k

   statistical dependency between the extracted Word2Vec-

   Where typically:

   N(tk) = lowercase(tk), tk V (6)

   Thus, the final preprocessed document representation

   based features and the sentiment labels [46]. Let X be a feature, and Y be the sentiment class label. The MI between X and Y is given by:

   MI(X, Y) = L I p(x, y)log ( p(x,y) ) (12)

   became:

   xX

   yY

   p(x)p(y)

   l 2 L

   di = {t, t, , t} (7)

   Where tdenotes tokens after sequential application of tokenization, stop-word removal, lemmatization, stemming, and normalization. This refined textual corpus ensured reduced noise, minimized dimensional redundancy, and improved semantic consistency before Word2Vec-based feature extraction and subsequent model training.

3. Word2Vec-Based Feature Extraction

i

Word2Vec is a method of embedding neural networks that can embed discrete tokens within text as continuous vectors, preserving their semantic and syntactic characteristics [43,44]. This study utilizes Word2Vec on the cleaned sequence of tokens Tclean = {wil, wi2, , win} to produce

semantic vectors suitable for modeling by Bi-LSTMs. Using

the given phrase as input, this algorithm seeks to generate context words with the highest possible probability. The equation corresponding to this optimization problem is:

N

Here, (x) and p(y) denote the marginal probability distributions, while p(x, y) denotes the joint probability distribution of feature X and class label Y. Each feature set

{Xl, X2, , Xm}, has its MI score calculated separately:

MI(Xk, Y), k = 1,2, , m (13)

The top-K features with the highest MI scores are selected:

selected = TopK(MI(Xk, Y)) (14)

This helps to ensure that only features with high dependency on the sentiment labels are selected for training the model. Removing irrelevant features and features with low dependency decreases the complexity of the classification model.

1. Data Splitting

   Splitting the dataset allows authors to evaluate the model’s

   max l I log P( w

   w ) (8)

   ability to generalize after feature extraction and selection. To

   0 N

   t=l

   -csjsc

   j1=0

   t+j t

   prevent models from being overfit to training data and to conduct fair performance comparisons, the suggested approach divides the processed dataset D into subsets (test

   and train). To keep the class distribution intact, the dataset is separated using stratified sampling:

   Where the query matrix is Q, K is the key matrix, V is the value matrix, dk for dimensionality of key vectors, and QKT

   ‘Dtrain

   ‘Dtest

   = ‘D, ‘Dtrain

   ‘Dtest

   = (15)

   present similarity score matrix. The contextual output at layer

   is:

   Typically,

   ‘Dtrain = 0.8N, ‘Dtest = 0.2N (16)

   This is a way to make sure that the performance evaluation is unbiased and there is no data leakage.

2. Model Designing and Training

   After the data is split, the three DL architectures are built and trained under the same experimental conditions to allow a fair comparison.

   1. Bi-LSTM Model

      The Bi-LSTM can learn contextual dependencies by taking both forward and backward motions of the text sequence into account [47]. In the proposed approach, it evaluates the effectiveness of recurrent architectures in handling informal and domain-variant social media text. For the input sequence Ei = {el, e2, , eT}:

      t f t t-l

      — = TM (e , ——–) (17)

      HL = Transformer(HL-l) (22)

      The contextual representation of the special [CLS] token is used for classification:

      [CLS]

      CLS = HL (23)

      Final prediction:

      y = Softmax(WCLS + b) (24)

      Where W is a subset of C×d, where d is the hidden size of the transformer. On the training dataset, the pretrained BERT model is fine-tuned with the help of cross-entropy loss. Fig. 4 shows the BERT model’s architecture.

      3) RoBERTa Model

      RoBERTa is an improvement over BERT that uses dynamic masking during pretraining and does away with the Next Sentence Prediction aim [51]. In this framework, RoBERTa evaluates whether enhanced pretraining

      techniques and dynamic masking strategies lead to improved

      — = TM (e , ——–) (18)

      t b t t+l

      generalization and robustness across diverse social media domains. It serves as an advanced comparative model to

      t t

      t = [—; —]; {where [; ] denotes concatenation} (19)

      Final sentiment prediction: y = Softmax(WT + b) (20)

      t

      t

      Where T is the sequence length (number of tokens), et M is embedding vector of token at position t, and d embedding dimension. t is the combined hidden state, — indicates the forward hidden state at time t, and — is the backward hidden state at time t. y C represents the predicted probability distribution, h stands for the number of LSTM hidden units, b is the bias vector, and W is a weight matrix in C×2h. The Bi-LSTM model’s design is shown in Fig. 3.

   2. BERT Model

   BERT employs multi-head self-attention to model deep contextual relationships [49]. Within the methodology, BERT is fine-tuned on the sentiment dataset to leverage its large-scale pretrained language knowledge. Its role is to evaluate the impact of deep contextual embeddings and attention mechanisms in improving sentiment classification accuracy, especially in handling contextual ambiguity and polysemy.

   Input embedding: E = Et0ken + Ep0siti0n + Esegment

   Where Et0ken is a token embedding matrix, Ep0siti0n shows positional embedding, and Esegment presents segment embedding. Self-attention mechanism:

   assess performance gains over both recurrent (Bi-LSTM) and baseline transformer (BERT) architectures. The architecture follows the same transformer encoder structure:

   Hl = Transformer(Hl-l) (25)

   The classification is performed using the contextual embedding of the first token:

   Attention(Q, K, V) = Softmax

   T

   QK

   ( )V (21)

   • dk

     yt-l

     yt

     yt+l

     (J

     (J

     (J

     –

     t-l

     —

     t

     –

     t+l

     Forward Pass

     LSTM

     LSTM

     LSTM

     ——–

     t-l

     —

     t

     ——–

     t+l

     Backward Pass

     LSTM

     LSTM

     LSTM

     xt-l

     xt

     xt+l

     Fig. 3. Architecture of Bi-LSTM [48]

     Tok1 Tok2 [Tok3] Tok4 Tok5

     Transformer Encoding Layers

     TE Layer 12

     TE Layer 11

     ……….

     TE Layer 10

     Add and Norm

     Feed Forward

     TE Layer 2

     TE Layer 1

     Encoding

     Tok1 Tok2 [Mask] Tok4 Tok5

     Add and Norm

     Multi-Head Attention

     INPUTS

     A Transformer Encoder Layers

     Fig. 4. BERT Architecture [50]

     y = Softmax(W0 + b) (26)

     RoBERTa is fine-tuned similarly using supervised cross- entropy loss. All models are trained by minimizing categorical cross-entropy loss. Parameter optimization is performed using the Adam optimizer:

3. SA and Model Evaluation

   SA is the automated process of figuring out what a piece of text means and putting it into one of three categories based on the emotional tone of the content: positive, negative, or neutral [53]. In this paper, SA is structured as a supervised multiclass classification problem, with each tweet

   0t+l

   = 0t

   m

   t

   T

   • vt+E

   (27)

   categorized into one of the established sentiment categories. Following feature extraction and preprocessing, the trained model f0 is applied to the input representation Xi, producing

   Where T presents the learning rate, and the first and second

   moment estimates are presented by mt, vt, respectively. Fig. 5 presents the architecture of the RoBERTa Model.

   This structured training and evaluation framework enables a fair comparative analysis of sequential modeling (Bi- LSTM) and transformer-based contextual modeling (BERT and RoBERTa) in multi-domain sentiment classification.

   a probability distribution across sentiment classes. The last

   projected label is the one that has the greatest chance of being correct. As categorical cross-entropy loss is minimized, optimizing model parameters ensures that the features related to sentiment are learned in a discriminative manner.

   Class Lebel

   .

   .

   RoBERTa

   .

   EN

   [CLS]

   T 1

   T 2

   .

   T N

   Solo Sentence

   E2

   E1

   E[CLS]

   E2

   E1

   E[CLS]

   EN

Fig. 5. RoBERTa Architecture [52]

To evaluate the proposed model, many metrics are employed. This captures differences in performance and considers all aspects of performance. Metrics are used to create an assessment of the model that includes the aspects of performance and efficiency. This assessment includes training time, inference time, and the parameter size. For this reason, the assessment includes an accurate and thorough comparison of the architectures of sequential models (Bi- LSTM) and transformer-based models (BERT, RoBERTa) in terms of accuracy and efficiency.

1. Continuous Monitoring

   Continuous monitoring, as defined in the proposed methodology, is the assessment of the deployed sentiment classification in terms of accuracy and robustness with respect to real-world usage. After deployment, the predictive accuracy of the classification model is evaluated using the streams of data that are input to the model to assess if there is degradation of accuracy, if there are imbalances, or if there are changes in the behavior of the model. Also, changes in the

   13: Compute MI between features and labels 14: Rank features based on MI score

   .

   15: Select top-k informative features 16: // Data Splitting

   .

   17: Split the dataset for Train and Test 18: // Model Designing

   .

   19: Initialize Bi-LSTM/BERT model 20: Define loss function (Cross-Entropy) 21: Define optimizer (Adam)

   22: // Model Training 23: For epoch = 1 to E:

   24: Forward propagation

   25: Compute loss

   26: Backpropagation

   27: Update model parameters 28: End For

   29: // Evaluation

   30: Predict sentiment on test data

   31: Compute Accuracy, Precision, Recall, F1-score, AUC 32: // Continuous Monitoring

   33: Monitor model performance on new data

   34: If performance degrades or a drift is detected: 35: Retrain model

   36: End If

   Return

   End Algorithm

   1. RESULTS AND DISCUSSION

      The results obtained from the experiments conducted in relation to the proposed comparative SA framework are given in this section to examine the efficiency of the Bi-LSTM, BERT, and RoBERTa models for the Sentiment140 dataset. The evaluation of the performance of the proposed approach is based on the capacity of the models to effectively perform sentiment classification while exhibiting robustness and efficiency. Performance evaluation is conducted using standard metrics to examine the classification performance of the models. The following are the metrics used:

      behavior of the model are evaluated by comparing the shifting of the statistical distributions data of the input streams with the data of the models training. A degradation in the models behavior and a shift in the statistical distribution are evaluated and adjusted to stabilize the proposed system.

2. Pseudo Code

Pseudo Code: Multimodal SA Framework

Input: Raw tweets (T)

Output: Predicted Sentiment Class ()

F1-score = 2(PrecisionRecall)

Precision+Recall

Accuracy = TP+TN

TP+TN+FP+FN

Precision = TP

TP+FP

Recall = TP

TP+FN

(28)

(29)

(30)

(31)

1: Collect dataset (T)

2: // Text Pre-processing

3: For each document d in D:

4: Perform Tokenization 5: Remove Stop Words

6: Apply Lemmatization/Stemming

7: Apply Normalization

8: End For

9: // Feature Extraction

10: Train Word2Vec model on processed corpus 11: Generate embedding vectors for each document 12: // Feature Selection

Table II illustrates the optimal configuration of hyperparameters used for the training of the Bi-LSTM, BERT, and RoBERTa models, which is important in conducting an experiment and comparing the performance of different models of sentiment classification.

Component Hyperparameter Value Used

TABLE II: HYPERPARAMETER SETTINGS USED FOR MODEL TRAINING

Dataset split

Word2Vec

Bi-LSTM

Train-test ratio 80:20:00 Sampling method Stratified sampling

Random state 42

Vector size 100

Window size 5

Minimum word count 2

Training algorithm Skip-gram Epochs 5

LSTM units 64 bidirectional units

Dense layer 32 units, ReLU

Output layer 2 units, Softmax

Loss function Sparse categorical cross- entropy

Optimizer Adam

Batch size 128

Epochs 1

tweets averaged 12.12 tokens, as shown in Table IV. The maximum token lengths for the negative and positive tweets are 32 and 30, respectively. These tokens supply the text for the following preprocessing steps.

TABLE IV: TOKENIZATION RESULTS

Sentime nt Class

Total Averag

Tweet e

s

Minimu Maximu Toke m m n

Tokens Tokens Tokens Coun

t

Maximum sequence

length 96

BERT

Batch size 8

Epochs 1

Fine-tuning sample size 400 tweets

Negative 10,00

0

Positive 10,00

0

12.95 0 32 9

12.12 0 30 8

Maximum sequence

length 96

RoBERTa

Batch size 8

Epochs 1

Fine-tuning sample size 400 tweets

Further, stop-word removal is executed to eliminate non- informative, high-frequency words. Fig. 6 shows that the mean token count drops for negative tweets from 12.95 to 8.6, and for positive tweets from 12.12 to 8.1, demonstrating a reduction in text redundancy with maintenance of the words

As part of the experiment, the Sentiment140 data set has been used as the input data set for testing the proposed SA models. To achieve better prediction results, a balanced sample containing 20,000 data points was selected from this source, where 50% of it consists of negative sentiment data and another 50% of positive sentiment data for 10,000 tweets each, as shown in Table III below. A balanced data distribution helps to minimize prediction biases and guarantees an accurate assessment of the performance of the utilized models.

Sentiment Label

Sentiment Class

Number of Percentage Records

Table III: Input Dataset Class Distribuion

0	Negative	10,000	50.0%
4	Positive	10,000	50.0%

that carry sentiment.

Fig. 6. Stop Word Removal Result

Following stop-word removal, lemmatization is performed, which modifies words to their base forms as they would be found in a dictionary and thereby preserves their meanings. This process decreases the variety of words while keeping the context intact. This process, along with retention of the sentiment words, also reduces the lexis used, which is presented in Table V.

Total Balanced Dataset

20,000 100%

TABLE V: ANALYSIS OF LEMMATIZATION

Original Tok	en After Stemmi	ng After Lemmatization
im	im	im
have	have	have
just	just	just
up	up	up
good	good	good
now	now	now
day	day	day
work	work	work

After a dataset is provided, preprocessing is performed on this dataset. The first step is tokenization, which breaks a tweet down into its constituent parts for processing with certain DL models. After balancing to a total of 20,000 tweets, negative tweets averaged 12.95 tokens while positive

all	all	all
get	get	get
out	out	out
like	like	like

Normalization of textual data through conversion of all characters into lowercahse letters, elimination of redundant

Negative

Negative

Positive

0.1

5

0.0

7

0.0

2

0.1

6

0.1

8

0.0

7

–

0.0

9

–

0.2

1

–

0.1

5

–

0.0

6

–

0.0

1

0.0

3

–

0.1

4

–

0.1

4

–

0.1

1

0.1

6

0.2

3

0.2

2

0.0

0

0.0

9

0.1

0

0.2

6

0.2

6

0.3

3

–

0.3

3

–

0.1

9

–

0.3

1

–

0.1

1

–

0.1

4

–

0.0

8

symbols, URLs, punctuation marks, and extra white spaces takes place next. As demonstrated in Fig. 7 below, the average number of characters in each tweet drops to about 50 from 75 for negative tweets and 76 from 48 for positive tweets, resulting in more streamlined texts for further processing through the Word2Vec model.

Fig. 7. Normalization Results

Following the preprocessing process, Word2Vec is used to represent the preprocessed tweets in vector form. In the Word2Vec model, an embedding space of dimension 100 is created using 8,386 vocabulary words, thereby generating the tweet embedding space of dimension (20,000 x 100) as presented in Table VI. This feature is used as the input for Bi- LSTM, BERT, and RoBERTa models to classify sentiments.

F1-

Trainin

Model Accuracy

Precis ion

Recall Scor AUC g Time

e

(s)

TABLE VI: WORD2VEC FEATURE EXTRACTION SUMMARY

MI is used after feature extraction to find the most useful Word2Vec embedding features for sentiment classification. MI analyzes the relationship between embedding features and sentiment labels; thus, features with the most discriminative power are chosen. As seen in Fig. 8, the features with the 20 highest MI values are chosen. Feature 74 has the highest MI value, at about 0.046. Feature 55 and Feature 63 are valued at

0.035 and 0.033, respectively. The feature selection process decreases redundant information, simplifies the models to lessen the computational time, and makes the Bi-LSTM, BERT, and RoBERTa models more effective.

Fourteen thousand tweets make up the training set, 70% of the entire set, while the remaining 30%, or 6,000 tweets, make up the test set. The training set is used to build the three models. The test set is used to assess the performance of each model objectively. The effectiveness of each model is assessed by measuring the accuracy, precision, recall, F1 score, AUC, and the time taken to train the model.

Table VIII provides a summary of the results. The RoBERTa model provides the best performance and classifies tweets with 98.56% accuracy and is trained in 5.46 seconds. The performance of the Bi-LSTM and BERT models is significantly worse and far less efficient than RoBERTa-based sentiment classification.

TABLE VIII: PERFORMANCE COMPARISON OF THE EVALUATED MODELS

Metric Value

Vector Size 100

Vocabulary Size 8,386

Tweet Embedding Matrix Shape (20,000, 100)

Bi- LSTM

63.68% 64.66

%

60.30

%

62.4

1%

68.45

%

125.62

Once the Word2Vec model is trained, each tweet gets transformed into an embedded vector of numbers. Table VII shows the first five sample tweet embeddings, where E-E

BERT 61.25% 100.0

0%

20.51

%

34.0

4%

75.11

%

56.44

are the first ten dimensions of the feature vector created, with a total of 100 dimensions. The continuous embeddings have

RoBER 98.56%

Ta

98.72

%

98.40

%

98.5

6%

99.21 5.46

%

captured the relationship between the words, and then they are fed to the Bi-LSTM, BERT, and RoBERTa models as the

input.

TABLE VII: SAMPLE WORD2VEC EMBEDDING VECTORS (FIRST 5 TWEETS)

Sentiment	E_	E_ E_		E_	E_	E_	E_	E_	E_	E_
Class	0	1 2		3	4	5	6	7	8	9
Positive	0.0 9	0.1 2	– 0.2 0	– 0.0 2	– 0.1 4	0.2 0	0.1 3	0.2 7	– 0.1 8	– 0.1 2
Negative	0.1 6	0.2 7	– 0.1 8	– 0.0 5	– 0.1 4	0.3 4	– 0.0 0	0.1 9	– 0.2 6	– 0.1 5

Fig. 8. Top Mutual Information Features

To examine the classification performance of the proposed framework further, confusion matrices were created for the Bi-LSTM, BERT, and RoBERTa models. See Figs. 9(a)(c). Confusion matrices show the correctly and incorrectly classified sentiment classes and support rfined analysis of model prediction behavior and classification performance.

The Bi-LSTM model classified 1,280 negative, 1,180 neutral, and 1,260 positive sentiments and showed high confusion for all sentiment classes, as shown in Fig. 9 (a). The BERT model classified negative sentiments the most, classifying 1,500 negative, 1,080 neutral, and 410 positive, showing bias toward the negative sentiment class as depicted in Fig. 9 (b). On the other hand, the proposed RoBERTa model showed significant classification performance, classifying 1,974 negative, 1,968 neutral, and 1,972 positive sentiments, with a few misclassifications for all three sentiment classes, as shown in Fig. 9 (c). The RoBERTa confusion matrix features most values aligned with the main diagonal compared to the other confusion matrices. It captures RoBERTas superior class distinction and choice of model, as it has the highest classification performance and the most classification models compared to the others.

The distribution of the sentiment prediction of each classification model is depicted in Table IX for the three sentiment classes. For instance, the Bi-LSTM model generated somewhat balanced predictions where 2,050 samples were classified as negative, 1,930 as neutral, and

(a)

(b)

(c)

Fig. 9. Confusion Matrices of All Three Models

TABLE IX: THREE-CLASS PREDICTION SUMMARY BASED ON CONFUSION MATRICES

Model Sentiment Record Count

2,020 as positive. On the other hand, BERT has an obvious prediction bias towards the negative class since it predicted

3,220 negative samples and only 860 positive samples. On the other hand, the proposed RoBERTa model managed to predict 2,000 samples in each category, making sure that the prediction is perfectly balanced. This is another indicator of the fact that the proposed RoBERTa model can accurately classify all three sentiment categories without any prediction biases.

Bi-LSTM

BERT

Negative 2050

Neutral 1930

Positive 2020

Negative 3220

Neutral 1920

Positive 860

RoBERTa Negative 2000

Neutral 2000

Positive 2000

1. Ablation Study

   Authors and Reference

   Year Dat aset

   Models Accurac y

   F1-

   Score

   To investigate the impact of each component, an ablation s

   study was undertaken for the proposed sentiment classification model with Word2Vec for feature extraction and MI for feature selection as shown in Table I0. The first step was to assess the baseline of Bi-LSTM, BERT, and

   Rahman 2025

   et al. [31]

   Elmitwal

   RoBERT

   a BiLSTM

   82.25%

   82.25

   %

   RoBERTa. Then the models were analyzed with Word2Vec Embedding and then with Word2Vec and MI. Among these models, the RoBERTa model, run without the assistance of

   li and

   Mehegan [35]

   2024 BERT 82.54% 81.14

   %

   Sen

   any other model, reached the baseline’s highest performance of 94.27%. The Bi-LSTM and BERT models were greatly surpassed by the performance of RoBERTa. The results indicate that the Word2Vec embedding was beneficial for all

   Tan et al.

   [39]

   Hafeez et al. [54]

   2023

   2026

   tim ent1 40

   RoBERT

   aLSTM

   DistilBE RT

   89.70% 90.00

   %

   86.00% 86.00

   %

   three models. After adding the Word2Vec embedding, RoBERTa reached a baseline of 96.48%. The highest performance was achieved by the proposed RoBERTa +

   Albladi et al. [55]

   2026 RoBERT

   a

   RoBER

   94.00% 94.10

   %

   Word2Vec + MI model with 98.56% accuracy, 98.72% precision, 98.40% recall, and 98.56% F1, all of which

   Propose d Model

   2026

   Ta + Word2V

   98.56% 98.56

   %

   indicate that the Word2Vec and MI models were highly successful.

   ec + MI

   Table X: ABLATION STUDY

   Model	Accurac y	Precisio Recall F1- n score
   BiLSTM only	58.92%	59.47% 57.86 58.65 % %
   BERT only	60.34%	61.12% 58.74 59.91 % %
   RoBerta only	94.27%	94.61% 93.82 94.21 % %
   BiLSTM+Word2Vec	61.18%	62.05% 59.84 60.92 % %
   BERT+Word2Vec	63.54%	64.38% 61.92 63.13 % %
   RoBerta+Word2Vec	96.48%	96.73% 96.21 96.47 % %
   BiLSTM+Word2Vec+ MI	63.68%	64.66% 60.30 62.41 % %
   BERT+Word2Vec+ MI	61.25%	100% 20.51 34.04 % %
   RoBerta+Word2Vec+ MI	98.56%	98.72% 98.40 98.56 % %

2. Comparative Analysis

Table I1 shows a comparison of the proposed sentiment classification framework and recently published work on the Sentiment140 dataset. It includes each study’s publication year, model used, accuracy, and F1 score. From the results, it can be reported that the proposed framework of RoBERTa + Word2Vec + Mutual Information achieved the best classification results at an accuracy of 98.56% with an F1 score of 98.56% at the time of this research, regarding the proposed classification framework.

TABLE XI: COMPARATIVE ANALYSIS OF PROPOSED MODEL WITH EXISTING STUDIES

It has been shown in these frameworks that combining the capabilities of a contextual transformer model with semantic feature learning and feature selection improves sentiment classification performance. This combination also enhances the robustness of the framework while addressing the limitations of sentiment classification observed in previous studies.

1. CONCLUSION AND FUTURE WORK

In this research, a sentiment analysis framework based on Bi-LSTM, BERT, and RoBERTa was presented for classifying social media sentiments using the Sentiment140 dataset. The designed framework included text preprocessing, Word2Vec feature extraction, MI feature selection, and DL based sentiment classification for enhancing prediction accuracy. The experiment results have confirmed that the proposed RoBERTa framework performed with better accuracy, with an accuracy of 98.56%, within only

5.46 s of training time, surpassing the performance of both Bi-LSTM and BERT regarding classification accuracy and computational efficiency. These results prove the effectiveness of improved context modeling and informative features to make sentiment classification more scalable. The novelty of this research lies in performing a systematic comparison of transformer-based and recurrent architectures under standardized experimental settings while incorporating an effective Word2Vec feature extraction and Mutual Information feature selection strategy. The future research may be dedicated to multilingual SA, code-mixed SA, the recognition of sarcasm and emotions, the development of methods for explainable artificial intelligence, and testing on large datasets of real-time social media texts.

REFERENCES

1. T. D. Francis, The Intersection of User-Generated Content and Social Media Disinformation. Media i Spoeczestwo, Vol. 22, No. 1/2, pp. 95 125, 2025

2. M. L. B. dos Santos, The so-called UGC: An updated definition of user-generated content in the age of social media, Online Information Review, ol. 46, no. 1, pp. 95113, 2022.

3. M. S. Alam, M. S. H. Mrida, and M. A. Rahman, Sentiment analysis in social media: How data science impacts public opinion knowledge integrates natural language processing (NLP) with artificial intelligence (AI), American Journal of Scholarly Research and Innovation, vol. 4, no. 1, pp. 63100, 2025.

4. P. R. Choudhary, Sentiment analysis of social media data for public opinion and market trend prediction, International Journal of Research Publications in Engineering, Technology and Management, vol. 8, no. 2, pp. 1176811771, 2025.

5. N. A. Sharma, A. S. Ali, and M. A. Kabir, A review of sentiment analysis: Tasks, applications, and deep learning techniques, International Journal of Data Science and Analytics, vol. 19, no. 3, pp. 351388, 2025.

6. M. Omar, A. Salah, and M. Mahdy, Sentiment analysis of social network contents using machine learning algorithms: A review, International Journal of Computers and Informatics, vol. 7, pp. 104125, 2025.

7. N. Nani, The influence of language on interaction and communication within social media platforms, Focus Journal: Language Review, vol. 2, no. 2, 2024.

8. A. D. Yacoub, S. Slim, and A. Aboutabl, A survey of sentiment analysis and sarcasm detection: Challenges, techniques, and trends, International Journal of Electrical and Computer Engineering Systems, vol. 15, no. 1, pp. 6978, 2024.

9. M. K. Nazir, C. N. Faisal, M. A. Habib, and H. Ahmad, Leveraging multilingual transformer for multiclass sentiment analysis in code-mixed data of low-resource languages, IEEE Access, vol. 13, pp. 75387554, 2025.

10. M. Muñoz-Organero and P. Queipo-Álvarez, Deep spatiotemporal model for COVID-19 forecasting, Sensors, vol. 22, no. 9, p. 3519, 2022.

11. Y. Zhang, L. Zhang, P. Pi, T. Li, Y. Chen, S. Peng, and Z. Ma, TARSS- Net: Temporal-Aware Radar Semantic Segmentation Network, Advances in Neural Information Processing Systems, vol. 37, pp. 49064933, 2024.

12. M. U. Salur and I. Aydin, A novel hybrid deep learning model for sentiment classification, IEEE Access, vol. 8, pp. 5808058093, 2020.

13. I. D. Mienye and Y. Sun, A survey of ensemble learning: Concepts, algorithms, applications, and prospects, IEEE Access, vol. 10, pp. 99129 99149, 2022.

14. S. Nazir, D. M. Dickson, and M. U. Akram, Survey of explainable artificial intelligence techniques for biomedical imaging with deep neural networks, Computers in Biology and Medicine, vol. 156, p. 106668, 2023.

15. O. Bashaddadh, N. Omar, M. Mohd, and M. N. A. Khalid, Machine learning and deep learning approaches for fake news detection: A systematic review of techniques, challenges, and advancements, IEEE Access, 2025.

16. S. K. Birthriya, P. Ahlawat, and A. K. Jain, Machine Learning-Based Smishing Detection Using Fuzzy Logic and TF-IDF Feature Engineering, Franklin Open, p. 100506, 2026.

17. Q. Li, H. Peng, J. Li, C. Xia, R. Yang, L. Sun, P. S. Yu, and L. He, A

    survey on text classification: From traditional to deep learning, ACM Transactions on Intelligent Systems and Technology, vol. 13, no. 2, pp. 1 41, 2022.

18. A. Chakhtouna, S. Sekkate, and A. Adib, Efficient bimodal emotion recognition system based on speech/text embeddings and ensemble learning fusion, Annals of Telecommunications, vol. 80, no. 5, pp. 379399, 2025.

19. M. Bordoloi and S. K. Biswas, Sentiment analysis: A survey on design framework, applications, and future scopes, Artificial Intelligence Review, vol. 56, no. 11, pp. 1250512560, 2023.

20. M. Wankhade, A. C. S. Rao, and C. Kulkarni, A survey on sentiment analysis methods, applications, and challenges, Artificial Intelligence Review, vol. 55, no. 7, pp. 57315780, 2022.

21. K. Alahmadi, S. Alharbi, J. Chen, and X. Wang, Generalizing sentiment analysis: A review of progress, challenges, and emerging directions, Social Network Analysis and Mining, vol. 15, no. 1, p. 45, 2025.

22. B. Omarov and Z. Zhumanov, Bidirectional long-short-term memory with attention mechanism for emotion analysis in textual content, International Journal of Advanced Computer Science and Applications, vol. 14, no. 6, 2023.

23. H. Alizadegan, B. R. Malki, A. Radmehr, H. Karimi, and M. A. Ilani, Comparative study of long short-term memory (LSTM), bidirectional LSTM, and traditional machine learning approaches for energy consumption prediction, Energy Exploration & Exploitation, vol. 43, no. 1, pp. 281301, 2025.

24. A. S. Mirdan, S. Buyrukolu, and M. R. Baker, Advanced deep learning techniques for sentiment analysis: Combining Bi-LSTM, CNN, and attention layers, International Journal of Advances in Intelligent Informatics, vol. 11, no. 1, 2025.

25. Y. Huang, J. Xu, J. Lai, Z. Jiang, T. Chen, Z. Li, Y. Yao, et al., Advancing transformer architecture in long-context large language models: A comprehensive survey, arXiv preprint arXiv:2311.12351, 2023.

26. N. Chen, T. Xu, M. Sun, C. Yao, and D. Yang, Understanding Video Transformers: A Review on Key Strategies for Feature Learning and Performance Optimization, Intelligent Computing, vol. 4, p. 0143, 2025.

27. S. Y. Baper and Z. K. Ismael, The impact of site contexts in increasing the diversity of architectural concepts: Using an interactive architectural approach, Buildings, vol. 14, no. 8, p. 2567, 2024.

28. C. Ling, X. Zhao, J. Lu, C. Deng, C. Zheng, J. Wang, T. Chowdhury, et al., Domain specialization as the key to make large language models disruptive: A comprehensive survey, ACM Computing Surveys, vol. 58, no. 3, pp. 139, 2025.

29. S. M. Sofi, A. Selamat, and Z. A. Shaffiei, An Adaptive Ensemble Machine Learning Classifier for Sentiment Analysis on Twitter, Journal of Advanced Research Design, vol. 136, no. 1, pp. 340357, 2026.

30. A. N. Arularasan, P. Ganeshkumar, T. Albalawi, and M. Alkhatib, A Radical Cycle GAN Based Term Memory Integrated Sentiment Analysis (CyG-TMSA) Framework for Social Media Applications, Cognitive Computation, vol. 18, no. 1, p. 1, 2026.

31. M. M. Rahman, A. I. Shiplu, Y. Watanobe, and M. A. Alam, RoBERTa-BiLSTM: A context-aware hybrid model for sentiment analysis, IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 9, no. 6, pp. 37883805, 2025.

32. A. G. Siddiqui, S. Ghulamani, and S. Hussain, A comparative analysis of text sentiment analysis algorithms using social media tweets, Pakistan Journal of Engineering and Technology, vol. 8, no. 3, pp. 1018, 2025.

33. J. Deng and Y. Liu, Research on sentiment analysis of online public opinion based on RoBERTaBiLSTMattention model, Applied Sciences, vol. 15, no. 4, p. 2148, 2025.

34. S. M. Karim, R. A. Rasul, and T. Sultana, Sentiment analysis of social media data for predicting consumer behavior trends using machine learning, arXiv preprint arXiv:2510.19656, 2025.

35. S. Elmitwalli and J. Mehegan, Sentiment analysis of COP9-related tweets: A comparative study of pre-trained models and traditional techniques, Frontiers in Big Data, vol. 7, p. 1357926, 2024.

36. L. He, Enhanced Twittr sentiment analysis with dual joint classifier integrating RoBERTa and BERT architectures, Frontiers in Physics, vol. 12, p. 1477714, 2024.

37. J. Tang, Leveraging transformer neural networks for enhanced sentiment analysis on online platform comments, in Proceedings of the 2023 4th International Conference on Machine Learning and Computer Application, pp. 256260, 2023.

38. M. Arbane, R. Benlamri, Y. Brik, and A. D. Alahmar, Social media- based COVID-19 sentiment classification model using Bi-LSTM, Expert Systems with Applications, vol. 212, p. 118710, 2023.

39. K. L. Tan, C. P. Lee, K. S. M. Anbananthen, and K. M. Lim, RoBERTa-LSTM: A hybrid model for sentiment analysis with transformer and recurrent neural network, IEEE Access, vol. 10, pp. 2151721525, 2022.

40. M. M. KazAnova, Sentiment140 dataset with 1.6 million tweets, Kaggle, 2017. [Online]. Available: https://www.kaggle.com/datasets/kazanova/sentiment140.

41. N. Garg and K. Sharma, Text preprocessing of multilingual data for sentiment analysis based on social network data, International Journal of Electrical & Computer Engineering, vol. 12, no. 1, pp. 776784, 2022.

42. S. Manda and B. V. Vardhan, A Systematic Literature Review on Sentimental Analysis using Machine Learning for Preprocessing, Feature Extraction and Classification, in Proceedings of the 3rd International Conference on Optimization Techniques in the Field of Engineering (ICOFE-2024), 2024.

43. A. Pak, A. Ziyaden, T. Saparov, I. Akhmetov, and A. Gelbukh, Word embeddings: A comprehensive survey, Computación y Sistemas, vol. 28, no. 4, pp. 20052029, 2024.

44. D. S. Asudani, N. K. Nagwani, and P. Singh, Impact of word embedding models on text analytics in a deep learning environment: A review, Artificial Intelligence Review, vol. 56, no. 9, pp. 1034510425, 2023.

45. M. Büyükkeçeci and M. C. Okur, A comprehensive review of feature selection and feature selection stability in machine learning, Gazi University Journal of Science, vol. 36, no. 4, pp. 15061520, 2023.

46. L. Huang, X. Zhou, L. Shi, and L. Gong, Time series feature selection method based on mutual information, Applied Sciences, vol. 14, no. 5, p. 1960, 2024.

47. J. Blake, A. S. M. Miah, K. Kredens, and J. Shin, Detection of AI- generated texts: A Bi-LSTM and attention-based approach, IEEE Access, 2025.

48. D. Naik and C. D. Jaidhar, A novel Multi-Layer Attention Framework for visual description prediction using bidirectional LSTM, Journal of Big Data, vol. 9, no. 1, p. 104, 2022.

49. G. Chen, Z. Cheng, Q. Lu, W. Weng, and W. Yang, Named Entity Recognition of Hazardous Chemical Risk Information Based on Multihead Self-Attention Mechanism and BERT, Wireless Communications and Mobile Computing, vol. 2022, no. 1, p. 8300672, 2022.

50. A. Noorian, A. Harounabadi, and M. Hazratifard, A sequential neural recommendation system exploiting BERT and LSTM on social media posts, Complex & Intelligent Systems, vol. 10, no. 1, pp. 721744, 2024.

51. X. Shu, BERT and RoBERTa for sarcasm detection: Optimizing performance through advanced fine-tuning, Applied and Computational Engineering, vol. 97, pp. 111, 2024.

52. R. Khusuma, W. Maharani, and P. H. Gani, Personality detection on Twitter user with RoBERTa, Journal Media Informatika Budidarma, vol. 7, no. 1, pp. 542553, 2023.

53. L. P. Hung and S. Alias, Beyond sentiment analysis: A review of recent trends in text based sentiment analysis and emotion detection, Journal of Advanced Computational Intelligence and Intelligent Informatics, vol. 27, no. 1, pp. 8495, 2023.

54. F. Hafeez, M. Hafeez, M. S. B. Imran, A. Qasim, N. Hussain, F. Ahmad, and G. Sidorov, A Comparative Study of Deep Learning and Transformer Models for Twitter Sentiment Analysis, International Journal of Combinatorial Optimization Problems & Informatics, vol. 17, no. 1, p. 85, 2026.

55. A. Albladi, M. K. Uddin, M. Islam, and C. Seals, TWSSenti: A Novel Hybrid Framework for Topic-Wise Sentiment Analysis on Social Media Using Transformer Models, arXiv preprint arXiv:2504.09896, 2025.