Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages

Bharathi Raja Chakravarthi, Ruba Priyadharshini, Anand Kumar Madasamy, Sajeetha Thavareesan, Elizabeth Sherly, Saranya Rajiakodi, Balasubramanian Palani, Malliga Subramanian, Subalalitha Cn, Dhivya ChinnappaNAACL 2025

TL;DR

This document presents the proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages (DravidianLangTech 2025). It showcases a diverse range of research addressing the technological marginalization of Dravidian languages, aiming to foster inclusive language technologies and ensure equitable access to digital information. The proceedings highlight numerous shared tasks and individual contributions focused on various NLP applications, from fake news detection to sentiment analysis and hate speech detection, across multiple Dravidian languages.

What Problem Does This Paper Address?

The core problem addressed by the DravidianLangTech 2025 workshop and its proceedings is the significant technological gap and underrepresentation of Dravidian languages in the fields of speech processing, computer vision, and natural language processing (NLP). While many dominant languages have successfully transitioned into the digital era, numerous regional and low-resource languages, including those from the Dravidian family, face substantial challenges in developing and deploying effective digital tools. This marginalization hinders equitable access to information and communication, and risks the digital extinction of these historically and culturally significant languages. The workshop series aims to bridge this gap by fostering research, development, and collaboration specifically for these languages.

Methods

The proceedings do not detail specific methods for a single paper but rather present a collection of research papers, each employing various methodologies. These include:

  • Transformer-based Models: Many papers utilize transformer architectures (e.g., BERT, XLM-RoBERTa, MuRIL) for tasks like text classification, sentiment analysis, and fake news detection.
  • Multimodal Approaches: Several works combine different modalities (text, audio, vision) to improve performance on tasks such as hate speech detection and meme analysis.
  • Ensemble Methods: Combining multiple models or techniques is used to enhance robustness and accuracy.
  • Custom MLP Architectures: Some papers propose novel neural network architectures tailored for specific tasks.
  • Traditional Machine Learning: While transformers are prevalent, some studies also explore traditional methods like Logistic Regression and SVMs, often as baselines or for comparison.
  • Feature Fusion Techniques: Various methods are employed to combine features from different sources or models.
  • Pre-trained Models: Leveraging pre-trained language models is a common strategy to overcome data scarcity.

Datasets

The proceedings do not focus on a single dataset but rather on research conducted across various Dravidian languages. Specific datasets mentioned or implied include:

  • Tamil Simile Dataset: A dataset specifically curated for computational processing of Tamil similes.
  • Tamil Political Comments: Datasets used for sentiment analysis of political discourse in Tamil.
  • Dravidian Languages: General datasets for tasks like fake news detection, hate speech detection, sentiment analysis, and abusive language detection across various Dravidian languages.
  • Malayalam and Tulu: Datasets for sentiment analysis and other NLP tasks in these languages.
  • Code-Mixed Text: Datasets containing text that mixes multiple languages, common in Dravidian language contexts.
  • YouTube Comments: Datasets derived from YouTube comments for analyzing abusive language.
  • Tamil and Malayalam Memes: Datasets for misogyny meme detection, often involving both image and text.

The proceedings also highlight several shared tasks organized as part of the workshop, which likely involved the creation or utilization of specific benchmark datasets for tasks such as:

  • Multimodal Hate Speech Detection in Dravidian languages
  • Detecting AI Generated Product Reviews in Dravidian Languages
  • Abusive Tamil and Malayalam Text targeting Women on Social Media
  • Misogyny Meme Detection
  • Sentiment Analysis in Tamil and Tulu

Key Insights and Ideas

  • The critical need for Dravidian language technologies: The sheer volume of papers and shared tasks underscores the urgent need to develop NLP tools for these languages, which have been historically underserved.
  • Multimodality is key: For tasks like hate speech and meme detection, combining text with other modalities (like images or audio) significantly improves performance, reflecting the complex nature of online content.
  • Transformer dominance, but not exclusivity: While transformers are the go-to architecture, research also explores their limitations and combinations with other methods, and sometimes traditional ML still holds its ground.
  • Addressing low-resource challenges: Many papers explicitly tackle the challenges of limited data by employing techniques like transfer learning, data augmentation, and leveraging pre-trained models.
  • The importance of shared tasks: The multiple shared tasks demonstrate a collaborative effort to standardize evaluation and drive progress across specific NLP challenges for Dravidian languages.
  • Focus on societal impact: A significant portion of the research addresses critical societal issues like fake news, abusive language, and misogyny, highlighting the potential of NLP to create safer and more equitable digital spaces.

Key Results

The proceedings present a wide array of results from numerous papers and shared tasks. Specific quantitative results are not aggregated here as this is a collection of proceedings. However, the general trend indicates:

  • High acceptance rates for shared tasks: This suggests active participation and a strong interest in these specific research areas.
  • Competitive performance across various NLP tasks: Papers report achieving state-of-the-art or near state-of-the-art results on benchmarks for tasks like fake news detection, sentiment analysis, and hate speech detection in Dravidian languages.
  • Demonstrated effectiveness of multimodal approaches: Many papers show superior performance when integrating multiple data modalities compared to text-only methods.
  • Improvements through fine-tuning and specialized architectures: Research on fine-tuning large language models and developing custom architectures often leads to significant gains in accuracy for specific tasks and languages.

Limitations and Future Work

  • Low-resource nature of languages: A recurring limitation is the scarcity of labeled data for many Dravidian languages, which necessitates innovative approaches for model training and generalization.
  • Domain specificity: Models trained on one domain (e.g., political tweets) may not perform well on another (e.g., product reviews), highlighting the need for more robust and adaptable models.
  • Ethical considerations: Detecting abusive content and fake news raises ethical concerns regarding bias, fairness, and potential misuse of technology.
  • Future Work:
    • Development of more comprehensive and diverse datasets for Dravidian languages.
  • Exploration of more efficient and interpretable models.
  • Cross-lingual transfer learning and zero-shot learning for under-resourced languages.
  • Addressing the ethical implications of NLP technologies.
  • Expanding research to include more Dravidian languages and dialects.
  • Further integration of speech and vision modalities with language processing.

Summary updated

BibTeX
@inproceedings{chakravarthi2025proceedings,
  author = {Bharathi Raja Chakravarthi and Ruba Priyadharshini and Anand Kumar Madasamy and Sajeetha Thavareesan and Elizabeth Sherly and Saranya Rajiakodi and Balasubramanian Palani and Malliga Subramanian and Subalalitha Cn and Dhivya Chinnappa},
  title = {{Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure