Automatic Bangla Headline Generation Using an Attention-Based Sequence-to-Sequence Model

Online Bangla content is growing quickly, yet Bangla remains a low-resource language for natural language processing (NLP). This study builds an automatic Bangla headline generation system that produces a short headline from a longer Bangla text. We compiled a dataset of 1,027 Bangla texts from news articles, blogs and social media, each paired with a human-written headline, and applied cleaning, contraction expansion, tokenization and pre-trained Bengali Word2Vec embeddings. An LSTM/GRU-based encoder-decoder with attention was implemented in TensorFlow and trained for 150 epochs on Google Colab, reaching a training loss of 0.0006. Qualitative examples show headlines that keep the core meaning of the source text. No ROUGE or BLEU evaluation was run, so the results should be read as a feasibility study rather than a benchmark. We discuss the limits of the current system and outline a path toward Transformer based models such as BanglaT5.