This repository and notebook contains code for in-progress implementation/experiments on Abstractive Text Summarization using Sequence-to-sequence RNNs and Beyond paper.

1. Create conda environment

conda env create -f environment.yml --gpu

conda env create -f environment-cpu.yml --cpu

2. Activate environment

source activate abs-sum -- gpu

source activate abs-sum-cpu -- cpu

3. Install dependencies (PyTorch, Fastai, etc) via:

pip install -r requirements.txt

4. Download spacy english module

python -m spacy download en

The dataset used is a subset of the gigaword dataset and can be found here.

It contains 3,803,955 parallel source & target examples for training and 189,649 examples for validation.

After downloading, we created article-title pairs, saved in tabular datset format (.csv) and extracted a sample subset (80,000 for training & 20,000 for validation). This data preparation can be found here.

An example article-title pair looks like this:

source: the algerian cabinet chaired by president abdelaziz bouteflika on sunday adopted the #### finance bill predicated on an oil price of ## dollars a barrel and a growth rate of #.# percent , it was announced here .

target: algeria adopts #### finance bill with oil put at ## dollars a barrel

Experimenting on the complete dataset (3M) would take a really long time (also $$$$). So in order to train and experiment faster we use our sample subset of 80,000.

• model architecture supports LSTM & GRU (biLSTM-to-uniLSTM or biGRU-to-uniGRU)
• implements batch data processing
• implements attention mechanism (Bahdanau et al. & Luong et al.(global dot))
• implements scheduled sampling (teacher forcing)
• implements tied embeddings
• initializes encoder-decoder with pretrained vectors (glove.6B.200d)
• implements custom training callbacks (tensorboard visualization for PyTorch, save best model & log checkpoint)
• implements attention plots

• Implement additional linguistic features embeddings
• Implement generator-pointer switch and replace unknown words by selecting source token with the highest attention score.
• Implement large vocabulary trick
• Implement sentence level attention
• Implement beam search during inference
• implement rouge evaluation