Benchmarking Multi-Task Learning for Sentiment Analysis and Offensive Language Identification in Under-Resourced Dravidian Languages

TL;DR

This paper evaluates multi-task learning models for sentiment analysis and offensive language detection in under-resourced Dravidian languages, demonstrating improved performance and efficiency over single-task models on code-mixed YouTube comments.

Contribution

It introduces a multi-task learning framework for under-resourced Dravidian languages, showing its effectiveness and providing benchmark results for these tasks.

Findings

Multi-task learning outperforms single-task models in accuracy.

Best models achieved high weighted F1-scores for all three languages.

The framework is adaptable to other sequence classification problems.

Abstract

To obtain extensive annotated data for under-resourced languages is challenging, so in this research, we have investigated whether it is beneficial to train models using multi-task learning. Sentiment analysis and offensive language identification share similar discourse properties. The selection of these tasks is motivated by the lack of large labelled data for user-generated code-mixed datasets. This paper works on code-mixed YouTube comments for Tamil, Malayalam, and Kannada languages. Our framework is applicable to other sequence classification problems irrespective of the size of the datasets. Experiments show that our multi-task learning model can achieve high results compared with single-task learning while reducing the time and space constraints required to train the models on individual tasks. Analysis of fine-tuned models indicates the preference of multi-task learning over single-task learning resulting in a higher weighted F1-score on all three languages. We apply two multi-task learning approaches to three Dravidian languages: Kannada, Malayalam, and Tamil. Maximum scores on Kannada and Malayalam were achieved by mBERT subjected to cross-entropy loss and with an approach of hard parameter sharing. Best scores on Tamil was achieved by DistilBERT subjected to cross-entropy loss with soft parameter sharing as the architecture type. For the tasks of sentiment analysis and offensive language identification, the best-performing model scored a weighted F1-score of (66.8\% and 90.5\%), (59\% and 70\%), and (62.1\% and 75.3\%) for Kannada, Malayalam, and Tamil on sentiment analysis and offensive language identification, respectively. The data and approaches discussed in this paper are published in Github\footnote{\href{https://github.com/SiddhanthHegde/Dravidian-MTL-Benchmarking}{Dravidian-MTL-Benchmarking}}.