KCRL@DravidianLangTech 2025: Multi-View Feature Fusion with XLM-R for Tamil Political Sentiment Analysis

Md Ayon Mia, Fariha Haq, Md. Tanvir Ahammed Shawon, Golam Sarwar Md. Mursalin, Muhammad Ibrahim KhanNAACL 2025

TL;DR

This paper tackles the challenging task of classifying Tamil political tweets into seven distinct sentiment categories, a crucial step for understanding public opinion in low-resource languages. The authors propose a novel Multi-View Feature Fusion (MVFF) architecture leveraging XLM-R, which combines different feature extraction strategies to achieve state-of-the-art results on this complex task.

What Problem Does This Paper Address?

Social media has become a dominant force in shaping public discourse, making accurate sentiment analysis essential. However, analyzing political sentiment, especially in low-resource languages like Tamil, presents significant challenges. Traditional sentiment analysis often focuses on binary (positive/negative) classifications, but political discourse is far more nuanced, encompassing categories like "Substantiated," "Sarcastic," and "Opinionated." Furthermore, the informal and context-dependent nature of social media, coupled with the linguistic complexities of Tamil, makes this task particularly difficult. This paper aims to bridge this gap by developing a robust system for multiclass sentiment analysis of Tamil political tweets.

Methods

The core of this work is the Multi-View Feature Fusion (MVFF) architecture, built upon the powerful XLM-R (Cross-lingual Language Model - RoBERTa) transformer model. The MVFF approach ingeniously combines three distinct feature extraction strategies from XLM-R's output:

  • [CLS] Token Representation: This special token, often used for classification tasks, captures a global representation of the input text.
  • Mean Pooling: Averaging the contextualized embeddings of all tokens in the sequence provides a summary of the overall sentiment.
  • Attention Mechanism: A dedicated attention layer is employed to weigh the importance of different token representations, allowing the model to focus on the most salient parts of the tweet for sentiment determination.

These three "views" of the text are then concatenated and fed into a feed-forward layer with a softmax activation for the final classification into one of the seven sentiment categories. This fusion strategy aims to capture a richer and more comprehensive understanding of the tweet's sentiment than any single method alone.

The authors also explored traditional Machine Learning (ML) approaches (Logistic Regression, SVM, Random Forest, XGBoost) using TF-IDF features and various Deep Learning (DL) architectures (LSTM, BiLSTM, BiLSTM-CNN, BiGRU-CNN) as baselines for comparison.

Datasets

The research utilizes a newly curated dataset for the Political Multiclass Sentiment Analysis of Tamil tweets shared task. This dataset comprises:

  • Training Set: 4,352 Tamil tweets.
  • Validation Set: 544 Tamil tweets.
  • Test Set: 544 Tamil tweets.

The dataset is annotated with seven distinct sentiment categories: Substantiated, Sarcastic, Opinionated, Positive, Negative, Neutral, and None of the above. A significant class imbalance was observed, with the "Opinionated" class being the largest and "None" being the smallest. To mitigate this, the authors employed random oversampling for minority classes during training.

Key Insights and Ideas

  • The Power of Feature Fusion: Combining multiple feature extraction strategies ([CLS], mean pooling, and attention) from a powerful pre-trained model like XLM-R significantly enhances sentiment classification performance, especially for nuanced tasks.
  • XLM-R's Cross-Lingual Prowess: XLM-R proves to be a strong foundation for low-resource languages like Tamil, demonstrating its ability to capture complex linguistic patterns even without extensive language-specific fine-tuning.
  • Class Imbalance is a Major Hurdle: The significant disparity in class distribution in the dataset poses a substantial challenge, requiring careful handling (like oversampling) to prevent model bias and ensure fair learning across all categories.
  • Context is King (and Hard to Capture): The paper highlights the difficulty in capturing broader conversational context and cultural references, which are crucial for accurate sentiment analysis, particularly in political discourse.

Key Results

The proposed Multi-View Feature Fusion (MVFF) with XLM-R and [CLS]+Mean+Attention pooling achieved the best overall performance on the test set, securing the 2nd position in the shared task. The model attained a macro-average F1-score of 0.37.

  • Transformer Models Outperform: Transformer-based approaches, particularly XLM-R with various pooling strategies, consistently outperformed traditional ML and standard DL architectures.
  • DL Models Show Promise: Deep learning models like LSTM and BiGRU+CNN showed competitive results (F1 scores of 0.35 and 0.34, respectively), indicating the benefit of recurrent and convolutional layers for feature extraction.
  • Confusion Matrix Insights: The confusion matrix reveals that the model excels at identifying "None of the above" (88%) and "Opinionated" (51%) categories. However, it struggles with distinguishing "Neutral" and "Negative" sentiments, often misclassifying them as "Opinionated." "Substantiated" content also shows dispersed misclassification, indicating challenges in identifying factual claims.

Limitations and Future Work

The authors acknowledge several limitations:

  • Modest F1-Scores: The overall F1-scores across all models are relatively modest, indicating inherent difficulties in Tamil political sentiment analysis.
  • Dataset Constraints: The limited size of the dataset and the class imbalance issues significantly impacted model development and performance.
  • Nuance and Context: The models struggled to capture nuanced political sentiments and often failed to consider broader conversational context or cultural references, which are vital for accurate classification.

Future work should focus on:

  • Addressing Class Imbalance: Exploring more advanced techniques for handling class imbalance.
  • Larger Datasets: Developing and utilizing larger, more comprehensive Tamil political datasets.
  • Enhanced Architectures: Investigating more sophisticated architectures that can better incorporate contextual information and cultural nuances for improved sentiment understanding.

Summary updated

BibTeX
@inproceedings{mia2025kcrl,
  author = {Md Ayon Mia and Fariha Haq and Md. Tanvir Ahammed Shawon and Golam Sarwar Md. Mursalin and Muhammad Ibrahim Khan},
  title = {{KCRL@DravidianLangTech 2025: Multi-View Feature Fusion with XLM-R for Tamil Political Sentiment Analysis}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure