Abstractive Text Summarization using Sequence-to-sequence RNNs and Beyond

Encoder-decoder model with attention mechanism

Word embedding

We used the Glove pre-trained vectors to initialize the word embeddings.

Encoder

Bidirectional GRU-RNN.

Decoder

Unidirectional GRU-RNN, with beamsearch.

Attention Mechanism

We Used BahdanauAttention.

Data

Data included in our github is a reduced dataset extracted from the dataset available at harvardnlp/sent-summary.

Evaluation

We used the ROUGE metric, from the package py-rouge.

Requirements

• Python 3
• Tensorflow version 1.x
• pip install -r requirements.txt

if google colab every dependency is installed in the notebook.

Usage

To use our implementation you simply go to the notebook text_summarization_feats.ipynb. And run the cells.

References:

1. D. Bahdanau, K. Cho, Y. Bengio, Neural machinetranslation by jointly learning to align and trans-late, arXiv preprint arXiv:1409.0473.

2. C.-Y. Lin, ROUGE: A package for automatic eval-uation of summaries, in: Text SummarizationBranches Out, Association for Computational Lin-guistics, Barcelona, Spain, 2004, pp. 74–81.URLhttps://www.aclweb.org/anthology/W04-1013.

3. R. Nallapati, B. Zhou, C. Gulcehre, B. Xi-ang, et al., Abstractive text summarization us-ing sequence-to-sequence rnns and beyond, arXivpreprint arXiv:1602.06023