Graph Based Extractive Tamil Text Summarization Evaluating FastText Embeddings

Automatic Text Summarization represents one of the most imperative and challenging applications of Natural Language Processing (NLP). Text Summarization extracts key information from a long text. Text summarization is categorized as Abstractive and Extractive methods. Summarization of texts in low-resource Indian regional languages like Tamil faces challenges in handling rich morphology and Out-Of-Vocabulary (OOV) words. Hence, effective word embedding is crucial for text summarization that captures the latent semantic relationships between words. Existing word embedding methods, Word2Vec and GloVe, struggle to handle the morphological complexities of Tamil. The FastText embedding model has proven efficient in handling OOV words, and morphological variations. Hence this work aims to study the effectiveness of FastText embedding in enhancing the extractive Tamil text summarization task performance. The performance of FastText has been evaluated using three graph based ranking algorithms viz. Text Ranking, Lex Ranking, and HITS on a curated Tamil sports news dataset. The resulting ROUGE scores establish the effectiveness of FastText against the baseline TF-IDF embedding. Our findings highlight the potential of sub-word-level semantic representations in enhancing text summarization for morphologically rich languages like Tamil.