santhoshkolloju/Abstractive-Summarization-With-Transfer-Learning

Leverages the Texar library to build an encoder-decoder architecture where pretrained BERT weights initialize the encoder while the decoder trains from scratch, enabling faster convergence than sequential LSTM models. Implements parallel training via Transformer self-attention mechanisms and includes a Flask-based inference server for generating abstractive summaries via HTTP POST requests. Requires TFRecord preprocessing of input story-summary pairs and supports full configuration customization through a dedicated config file.

412 stars. No commits in the last 6 months.