coolrb / nlu-email-title

Natural Language Understanding (Text Summarization) task for autonomous email title generation.

NLU-email-title

Informal and noisy text summarization is a meaningful task with practical use potentials. In this project, we adopted the data-driven attention-based neural model on the task to generate email titles. The experiment evaluations show that our model achieves 21.22 Rouge-L and 17.10 BLEU. We showed that the attention-based model outperforms the extractive methods, though informal text summarization remains to be a challenge.

Datasets

The Enron Corpus were used in this project. The original dataset can be downloaded at https://www.cs.cmu.edu/~enron/. It is parsed by the Parsing Enron Data.ipynb under the notebook folder. The parsing outputs a CSV file that contains subject line and first email body.

Baseline Model

The baseline model is TextRank, which is a extractive graph-based model. The notebook baseline_test.ipynb demonstrated the process of extracting subject lines and computing rouge scores. The extractive technique does not require further training.

Encoder-Decoder Model

This is an abstractibe model, also known as ABS. We use notebook Training Encoder Decoder.ipynb to train the model. Our encoder is a bidirectional GRU (gated recurrent unit) and the decoder is attentional based. The training takes roughly 4 hours on the entire preprocessed enron corpus.