ymfa / SummaRuNNer

Implementation of SummaRuNNer (runnable on custom texts)

The PyTorch Implementation Of SummaRuNNer

Models

1. RNN_RNN

2. CNN_RNN

3. Hierarchical Attention Networks

Setup

Requires pipenv. Use pip install pipenv if not installed.

pipenv install
pipenv shell

Usage

# train
python main.py -device 0 -batch_size 32 -model RNN_RNN -seed 1 -save_dir checkpoints/XXX.pt
# test
python main.py -device 0 -batch_size 1 -test -load_dir checkpoints/XXX.pt

Summarize your own texts

For example, you have some documents some/folder/*/article.txt. The first step is to tokenize them and pack them into a JSON file. To do this, run:

python make_data.py "some/folder/*/article.txt" data/my_collection.json

When running main.py, you can use the new option -num_tok to control the exact number of words per each output summary, or the existing -topk option (select k sentences). If you're running PyTorch without CUDA, execute git apply no_cuda.patch first.

To untokenize the output summaries, use

python put_back_summaries.py outputs/hyp/ "some/folder/*/"

to send the untokenized summaries to the same directories as their originals.

Pretrained models

1. RNN_RNN(checkpoints/RNN_RNN_seed_1.pt)
2. CNN_RNN(checkpoints/CNN_RNN_seed_1.pt)
3. AttnRNN(checkpoints/AttnRNN_seed_1.pt)

Result

DailyMail(75 bytes)

model	ROUGE-1	ROUGE-2	ROUGE-L
SummaRNNer(Nallapati)	26.2	10.8	14.4
RNN-RNN	26.0	11.5	13.8
CNN-RNN	25.8	11.3	13.8
Hierarchical Attn Net	26.0	11.4	13.8

Blog

• 用PyTorch搭建抽取式摘要系统

Download Data:

• 百度云:https://pan.baidu.com/s/1LV3iuuH1NjxuAJd0iz14lA 密码:ivzl

• Google Driver:data.tar.gz

• Source Data:Neural Summarization by Extracting Sentences and Words

Evaluation

• Tools

Acknowledge

• Thanks for @AlJohri's contribution