1.
Enhancing Trust and Interpretability in Malayalam Sentiment Analysis with Explainable AI
Authors: Anitha R, Rajeev R R, Meharuniza Nazeem, and Navaneeth S
Published in: Proceedings of the 21st International Conference on Natural Language Processing
(ICON), December 2024
Summary: Sentiment analysis in languages such as Malayalam presents unique challenges
due to linguistic complexity and the lack of large annotated datasets. Anitha R et al. (2024) have
proposed an explainable AI (XAI)-based framework to enhance the transparency and reliability
of sentiment classification models for Malayalam. The study utilizes Support Vector Machines
(SVM) and Random Forest (RF) classifiers trained on the Kerala Political Opinion Corpus,
achieving an impressive accuracy of 85.07% with the RF classifier.
The integration of Local Interpretable Model-Agnostic Explanations (LIME) has significantly
improved user trust by providing understandable justifications for model predictions. LIME
works by approximating black-box machine learning models with interpretable surrogate
models, allowing users to comprehend why a specific sentiment was assigned to a given text.
The study also highlights the need for domain-adaptive sentiment lexicons and proposes a semi-
supervised learning approach to expand sentiment-labeled datasets in Malayalam.
Future research in this area could explore the use of transformer-based architectures like
mBERT and IndicBERT, which have shown promise in handling morphologically rich languages.
Additionally, crowd-sourced data annotation efforts can help expand sentiment datasets,
improving the generalizability of models across diverse Malayalam dialects.
2. Deep Learning-Based Sentiment Analysis for Social Media: A Focus on Multimodal and
Aspect-Based Approaches
Author: Bowen Feng
Published in: Applied and Computational Engineering, February 2024
Summary: The paper underscores the significance of sentiment analysis, commonly known as
opinion mining, which leverages deep learning systems to discern human emotions and
subjective sentiments across various subjects. This process involves intricate computational
models that analyze and interpret the emotional undertones embedded within text,
categorizing them as positive, negative, or neutral. The ability to quantify sentiment is
particularly vital in the digital age, where opinions shared on social media can greatly influence
public sentiment and consumer behavior. The study consolidates relevant research within the
field conducted in recent years and anticipates future trajectories and impacts of this rapidly
evolving domain.
The paper highlights the transformational effects of deep learning on sentiment analysis,
emphasizing how these advanced models have enhanced the accuracy and depth of emotion
recognition and interpretation.
While the abstract does not provide specific details on multimodal techniques, the title suggests
a focus on integrating multiple data types (e.g., text, images) to improve sentiment analysis
outcomes.
The study likely discusses aspect-based sentiment analysis, which involves examining
sentiments related to specific components or features of a subject, providing a more granular
understanding of public opinion.
3. Enhancing Sentiment Analysis on Social Media with Novel Preprocessing Techniques
Authors: Khouloud Safi Eljil, Farid Nait-Abdesselam, Essia Hamouda, and Mohamed Hamdi
Published in: Journal of Advances in Information Technology, November 2023
Summary: This research explores the enhancement of sentiment analysis on social media
platforms through novel preprocessing techniques. The authors compared conventional
machine learning approaches, such as Naive Bayes, Logistic Regression, and Support Vector
Machine, against Bidirectional Encoder Representations from Transformer (BERT). The study
introduced a new preprocessing technique that significantly improved the performance of
conventional models. BERT outperformed all models with an accuracy of approximately 94%,
albeit with higher computational costs. The research also highlighted the effectiveness of
combining unigram and bigram words for better understanding of negation and proposed
approaches for handling emoticons and emojis to enhance sentiment analysis and sarcasm
interpretation.
The research demonstrates that combining unigrams and bigrams in feature extraction provides
a more nuanced understanding of negation in text, surpassing the effectiveness of unigrams
alone.
Recognizing the importance of emoticons and emojis in conveying sentiment on social media,
the authors propose a preprocessing approach that replaces emoticons with corresponding
textual meanings using regular expressions.
These studies contribute to the advancement of sentiment analysis by exploring various
methodologies, languages, and applications, reflecting the diverse approaches in recent
research.
4. Enhanced Financial Sentiment Analysis and Trading Strategy Development Using Large
Language Models :
Author: Kemal Kirtac and Guido Germano (2024)
Published in : Proceedings of the 14th Workshop on Computational Approaches to Subjectivity,
Sentiment, & Social Media Analysis (WASSA 2024)
Summery: Among the LLMs evaluated—OPT (a GPT-3-based model), BERT, FinBERT, LLAMA 3,
and RoBERTa—the OPT model demonstrated the highest accuracy in predicting stock market
returns, achieving a 74.4% accuracy rate. This performance notably surpasses that of the
traditional Loughran-McDonald dictionary model, which recorded an accuracy of 50.1%.
Regression analyses revealed a robust positive correlation between the sentiment scores
generated by the OPT model and subsequent daily stock returns. Specifically, the OPT model's
scores had coefficients of 0.274 and 0.254 in different regression models, indicating a significant
predictive relationship.
A self-financing trading strategy based on the OPT model's sentiment scores achieved a Sharpe
ratio of 3.05 over the sample period. In comparison, strategies based on the Loughran-
McDonald dictionary model yielded a Sharpe ratio of 1.23, highlighting the enhanced
performance and potential profitability of strategies informed by advanced LLMs
[Link] Sentiment Analysis and Prediction of Stock Price Using FinBERT, GPT-4, and
Logistic Regression: A Data-Driven Approach
Author : Olamilekan Shobayo, Sidikat Adeyemi-Longe, Olusogo Popoola, and Bayode Ogunleye
(2024)
Summery :A finance-specific adaptation of the Bidirectional Encoder Representations from
Transformers (BERT) model, designed to understand the nuances of financial text.
The fourth iteration of the Generative Pre-trained Transformer, known for its versatility in
natural language processing tasks.A traditional statistical method employed for binary
classification tasks.
Logistic Regression: Demonstrated superior performance with an accuracy of 81.83%
and a ROC AUC of 89.76%, outperforming the more complex AI models.
FinBERT: Offered moderate performance but was resource-intensive, indicating a trade-
off between computational demands and analytical [Link]+1MDPI+1
GPT-4: Achieved an accuracy of 54.19%, showcasing potential in handling complex data
despite lower accuracy in this context
The study involved classifying market sentiment, generating sentiment scores, and
predicting market price movements. Model evaluation was conducted using metrics
such as accuracy, precision, recall, F1 score, and ROC AUC. Hyperparameter optimization
was performed using Optuna, and cross-validation techniques were applied to ensure
model robustness.