TFIDF to LLMs: A Comparative Analysis of Bengali Text Summarization Approaches and Techniques

Automatic Text Summarization is the process of condensing an input text into a concise output without losing any core information, the importance of which is increasing rapidly in this digital era. Text summarization is broadly divided into two approaches, i.e., extractive and abstractive. These approaches are implemented using many techniques, which are commonly grouped into five categories: statistical, graph-based, machine learning, deep learning, and transformer-based methods. To conduct a thorough investigation of the strengths and weaknesses of these techniques, we evaluated nine popular and reproducible text summarization models, including four Large Language Models (LLM), i.e., Gemma 2, LLaMa 3.3, Mistral, and DeepSeek-R1, based on their performance in Bengali text summarization. We found out that the abstractive summaries produced by LLMs like Gemma and especially LLaMa are very coherent and human-like, but they either sufer linguistically (Gemma) or factually (LLaMa). Extractive models, especially ML and DL models, are very competent at creating Bengali summaries that are contextually and factually correct but are less coherent. These trends are also seen in the overall comparison of extractive and abstractive methods. This study highlights the viability of a hybrid approach to create both coherent and factually correct summaries in low-resource languages by revealing the strengths and weaknesses of different techniques.