DHGN

This is the pytorch code for our CCL 2021 paper Incorporating Commonsense Knowledge into Abstractive Dialogue Summarization via Heterogeneous Graph Networks [arXiv Preprint].

Outputs

• Output summaries are available at summaries/summary_rerank.txt.

Requirements

• We use Conda python 3.7 and strongly recommend that you create a new environment: conda create -n dhgn python=3.7.
• Run the following command: pip install -r requirements.txt.

* pytorch_geometric

• We use pytorch_geometric for GNN implementation.
• Download *.whl files from https://pytorch-geometric.com/whl/torch-1.4.0.html
• Run following commands:

pip install torch_spline_conv-latest+cu100-cp37-cp37m-linux_x86_64.whl
pip install torch_cluster-latest+cu100-cp37-cp37m-linux_x86_64.whl
pip install torch_sparse-latest+cu100-cp37-cp37m-linux_x86_64.whl
pip install torch_scatter-latest+cu100-cp37-cp37m-linux_x86_64.whl
pip install torch-geometric==1.4.3

Reproduce Results

You can follow the following steps to reproduce the best results in our paper.

* download checkpoints

Download checkpoints here. Put the checkpoint samsum.pt under the project dir DHGN/models/samsum.pt.

* translate

We can get summaries/summary.txt

python translate.py \
-src data/test.src.txt \
-edge_index data/test.edge_index.txt \
-edge_type data/test.edge_type.txt \
-node_type data/test.node_type.txt \
-max_length 1000 -min_length 19 -beta 1 -ignore_when_blocking "." "," -stepwise_penalty -coverage_penalty summary -n_best 4 -batch_size 64 -beam_size 10 -dynamic_dict -share_vocab  -block_ngram_repeat 3 -replace_unk \
-model models/samsum.pt \
-output summaries/summary.txt \
-gpu X

* rerank

• python rerank.py -n 4 -c summaries/summary.txt

* test rouge score

• python py_rouge_test.py -c summaries/summary_rerank.txt
• Output >>> ROUGE 1-2-L F: 42.03-18.07-39.56

From Scratch

* preprocess

(1) Preprocess ICSI dataset.

python preprocess.py \
  -train_src data/train.src.txt \
  -train_edge_index data/train.edge_index.txt \
  -train_edge_type data/train.edge_type.txt \
  -train_node_type data/train.node_type.txt \
  -train_tgt data/train.tgt.txt \
  -valid_src data/valid.src.txt \
  -valid_edge_index data/valid.edge_index.txt \
  -valid_edge_type data/valid.edge_type.txt \
  -valid_node_type data/valid.node_type.txt \
  -valid_tgt data/valid.tgt.txt \
  -save_data data/samsum \
  -dynamic_dict \
  -share_vocab \
  -src_words_min_frequency 0 \
  -tgt_words_min_frequency 0 \
  -src_seq_length 10000 \
  -tgt_seq_length 10000 \
  -src_seq_length_trunc 10000 \
  -tgt_seq_length_trunc 10000 \
  -overwrite

(2) Create pre-trained word embeddings.

download glove.6B.100d.txt from here.

python embeddings_to_torch.py -emb_file_both data/glove.6B.100d.txt \
-dict_file data/samsum.vocab.pt \
-output_file data/samsum_embeddings

* train

CUDA_VISIBLE_DEVICES=x python train.py -config ./config/train.yml -save_config ./config/train.txt
CUDA_VISIBLE_DEVICES=x python train.py -config ./config/train.yml

* translate

python translate.py \
-src data/test.src.txt \
-edge_index data/test.edge_index.txt \
-edge_type data/test.edge_type.txt \
-node_type data/test.node_type.txt \
-max_length 1000 -min_length 19 -beta 1 -ignore_when_blocking "." "," -stepwise_penalty -coverage_penalty summary -n_best 4 -batch_size 64 -beam_size 10 -dynamic_dict -share_vocab  -block_ngram_repeat 3 -replace_unk \
-model models/xxxx.pt \
-output summaries/xxxx.txt \
-gpu X

* rerank

• python rerank.py -n 4 -c summaries/xxxx.txt

* test rouge score

• python py_rouge_test.py -c summaries/summary_rerank.txt