DDAMS

This is the pytorch code for our IJCAI 2021 paper Dialogue Discourse-Aware Graph Model and Data Augmentation for Meeting Summarization [Arxiv Preprint].

Update

2021.6.9 update pretrained models for AMI and ICSI. here, under the qg_pretrain dir;
2021.6.5 update Dialogue Discourse Parser;

Outputs

Output summaries are available at AMI and ICSI.

Requirements

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

Data

You can download data here, put the data under the project dir DDAMS/data/xxx.

data/ami
- data/ami/ami: preprocessed meeting data
- data/ami/ami_qg: pseudo summarization data.
- data/ami/ami_reference: golden reference for test file.
data/icsi
- data/icsi/icsi: preprocessed meeting data
- data/icsi/icsi_qg: pseudo summarization data.
- data/icsi/icsi_reference: golden reference for test file.
data/glove: pre-trained word embedding glove.6B.300d.txt.

Reproduce Results

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

download checkpoints

Download checkpoints here. Put the checkpoints, including AMI.pt and ICSI.pt, under the project dir DDAMS/models/xx.pt.

translate

Produce final summaries.

For AMI, we can get summaries/ami_summary.txt.

CUDA_VISIBLE_DEVICES=X python translate.py -batch_size 1 \
               -src data/ami/ami/test.src \
               -tgt data/ami/ami/test.tgt \
               -seg data/ami/ami/test.seg \
               -speaker data/ami/ami/test.speaker \
               -relation data/ami/ami/test.relation \
               -beam_size 10 \
               -share_vocab \
               -dynamic_dict \
               -replace_unk \
               -model models/AMI.pt \
               -output summaries/ami_summary.txt \
               -block_ngram_repeat 3 \
               -gpu 0 \
               -min_length 280 \
               -max_length 450

For ICSI, we can get summaries/icsi_summary.txt.

CUDA_VISIBLE_DEVICES=x python translate.py -batch_size 1 \
               -src data/icsi/icsi/test.src \
               -seg data/icsi/icsi/test.seg \
               -speaker data/icsi/icsi/test.speaker \
               -relation data/icsi/icsi/test.relation \
               -beam_size 10 \
               -share_vocab \
               -dynamic_dict \
               -replace_unk \
               -model models/ICSI.pt \
               -output summaries/icsi_summary.txt \
               -block_ngram_repeat 3 \
               -gpu 0 \
               -min_length 400 \
               -max_length 550

remove tags

<t> and </t> will raise errors for ROUGE test. So we should first remove them. (following OpenNMT)

sed -i 's/ <\/t>//g' summaries/ami_summary.txt
sed -i 's/<t> //g' summaries/ami_summary.txt
sed -i 's/ <\/t>//g' summaries/icsi_summary.txt
sed -i 's/<t> //g' summaries/icsi_summary.txt

test rouge score

Change pyrouge.Rouge155() to your local path.

Output format >> ROUGE(1/2/L): xx.xx-xx.xx-xx.xx

python test_rouge.py -c summaries/ami_summary.txt
python test_rouge_icsi.py -c summaries/icsi_summary.txt

ROUGE score

You will get following ROUGE scores.

	ROUGE-1	ROUGE-2	ROUGE-L
AMI	53.15	22.32	25.67
ICSI	40.41	11.02	19.18

From Scratch

Dialogue Discourse Parser

Please refer to DialogueDiscourseParser dir.

For AMI

Preprocess

(1) Preprocess AMI dataset.

python preprocess.py -train_src data/ami/ami/train.src \
                     -train_tgt data/ami/ami/train.tgt \
                     -train_seg data/ami/ami/train.seg \
                     -train_speaker data/ami/ami/train.speaker \
                     -train_relation data/ami/ami/train.relation \
                     -valid_src data/ami/ami/valid.src \
                     -valid_tgt data/ami/ami/valid.tgt \
                     -valid_seg data/ami/ami/valid.seg \
                     -valid_speaker data/ami/ami/valid.speaker \
                     -valid_relation data/ami/ami/valid.relation \
                     -save_data data/ami/AMI \
                     -dynamic_dict \
                     -share_vocab \
                     -lower \
                     -overwrite

(2) Create pre-trained word embeddings.

python embeddings_to_torch.py -emb_file_both data/glove/glove.6B.300d.txt \
-dict_file data/ami/AMI.vocab.pt \
-output_file data/ami/ami_embeddings

(3) Preprocess pseudo summarization dataset.

python preprocess.py -train_src data/ami/ami_qg/train.src \
                     -train_tgt data/ami/ami_qg/train.tgt \
                     -train_seg data/ami/ami_qg/train.seg \
                     -train_speaker data/ami/ami_qg/train.speaker \
                     -train_relation data/ami/ami_qg/train.relation \
                     -save_data data/ami/AMIQG \
                     -lower \
                     -overwrite \
                     -shard_size 500 \
                     -dynamic_dict \
                     -share_vocab

Train

(1) we first pre-train our DDAMS on the pseudo summarization dataset.

run the following command to save config file (-save_config).
remove -save_config and rerun the command to start the training process.

CUDA_VISIBLE_DEVICES=X python train.py -save_model ami_qg_pretrain/AMI_qg\
           -data data/ami/AMIQG \
           -speaker_type ami \
           -batch_size 64 \
           -learning_rate 0.001 \
           -share_embeddings \
           -share_decoder_embeddings \
           -copy_attn \
           -reuse_copy_attn \
           -report_every 30 \
           -encoder_type hier3 \
           -global_attention general \
           -save_checkpoint_steps 500 \
           -start_decay_steps 1500 \
           -pre_word_vecs_enc data/ami/ami_embeddings.enc.pt \
           -pre_word_vecs_dec data/ami/ami_embeddings.dec.pt \
           -log_file logs/ami_qg_pretrain.txt \
           -save_config logs/ami_qg_pretrain.txt

(2) fine-tuning on AMI.

CUDA_VISIBLE_DEVICES=X python train.py -save_model ami_final/AMI \
           -data data/ami/AMI \
           -speaker_type ami \
           -train_from ami_qg_pretrain/xxx.pt  \
           -reset_optim all \
           -batch_size 1 \
           -learning_rate 0.0005 \
           -share_embeddings \
           -share_decoder_embeddings \
           -copy_attn \
           -reuse_copy_attn \
           -encoder_type hier3 \
           -global_attention general \
           -dropout 0.5 \
           -attention_dropout 0.5 \
           -start_decay_steps 500 \
           -decay_steps 500 \
           -log_file logs/ami_final.txt \
           -save_config logs/ami_final.txt

Translate

CUDA_VISIBLE_DEVICES=X python translate.py -batch_size 1 \
               -src data/ami/ami/test.src \
               -tgt data/ami/ami/test.tgt \
               -seg data/ami/ami/test.seg \
               -speaker data/ami/ami/test.speaker \
               -relation data/ami/ami/test.relation \
               -beam_size 10 \
               -share_vocab \
               -dynamic_dict \
               -replace_unk \
               -model xxx.pt \
               -output xxx.txt \
               -block_ngram_repeat 3 \
               -gpu 0 \
               -min_length 280 \
               -max_length 450

For ICSI

Preprocess

(1) Preprocess ICSI dataset.

python preprocess.py -train_src data/icsi/icsi/train.src \
                     -train_tgt data/icsi/icsi/train.tgt \
                     -train_seg data/icsi/icsi/train.seg \
                     -train_speaker data/icsi/icsi/train.speaker \
                     -train_relation data/icsi/icsi/train.relation \
                     -valid_src data/icsi/icsi/valid.src \
                     -valid_tgt data/icsi/icsi/valid.tgt \
                     -valid_seg data/icsi/icsi/valid.seg \
                     -valid_speaker data/icsi/icsi/valid.speaker \
                     -valid_relation data/icsi/icsi/valid.relation \
                     -save_data data/icsi/ICSI \
                     -src_seq_length 20000 \
                     -src_seq_length_trunc 20000 \
                     -tgt_seq_length 700 \
                     -tgt_seq_length_trunc 700 \
                     -dynamic_dict \
                     -share_vocab \
                     -lower \
                     -overwrite

(2) Create pre-trained word embeddings.

python embeddings_to_torch.py -emb_file_both data/glove/glove.6B.300d.txt \
-dict_file data/icsi/ICSI.vocab.pt \
-output_file data/icsi/icsi_embeddings

(3) Preprocess pseudo summarization dataset.

python preprocess.py -train_src data/icsi/icsi_qg/train.src \
                     -train_tgt data/icsi/icsi_qg/train.tgt \
                     -train_seg data/icsi/icsi_qg/train.seg \
                     -train_speaker data/icsi/icsi_qg/train.speaker \
                     -train_relation data/icsi/icsi_qg/train.relation \
                     -save_data data/icsi/ICSIQG \
                     -lower \
                     -overwrite \
                     -shard_size 500 \
                     -dynamic_dict \
                     -share_vocab

Train

(1) pre-training.

CUDA_VISIBLE_DEVICES=X python train.py -save_model icsi_qg_pretrain/ICSI \
           -data data/icsi/ICSIQG \
           -speaker_type icsi \
           -batch_size 64 \
           -learning_rate 0.001 \
           -share_embeddings \
           -share_decoder_embeddings \
           -copy_attn \
           -reuse_copy_attn \
           -report_every 30 \
           -encoder_type hier3 \
           -global_attention general \
           -save_checkpoint_steps 500 \
           -start_decay_steps 1500 \
           -pre_word_vecs_enc data/icsi/icsi_embeddings.enc.pt \
           -pre_word_vecs_dec data/icsi/icsi_embeddings.dec.pt \
           -log_file logs/icsi_qg_pretrain.txt \
           -save_config logs/icsi_qg_pretrain.txt

(2) fine-tuning on ICSI.

CUDA_VISIBLE_DEVICES=X python train.py -save_model icsi_final/ICSI \
           -data data/icsi/ICSI \
           -speaker_type icsi \
           -train_from icsi_qg_pretrain/xxx.pt  \
           -reset_optim all \
           -batch_size 1 \
           -learning_rate 0.0005 \
           -share_embeddings \
           -share_decoder_embeddings \
           -copy_attn \
           -reuse_copy_attn \
           -encoder_type hier3 \
           -global_attention general \
           -dropout 0.5 \
           -attention_dropout 0.5 \
           -start_decay_steps 1000 \
           -decay_steps 100 \
           -save_checkpoint_steps 50 \
           -valid_steps 50 \
           -log_file logs/icsi_final.txt \
           -save_config logs/icsi_final.txt

Translate

CUDA_VISIBLE_DEVICES=x python translate.py -batch_size 1 \
               -src data/icsi/icsi/test.src \
               -seg data/icsi/icsi/test.seg \
               -speaker data/icsi/icsi/test.speaker \
               -relation data/icsi/icsi/test.relation \
               -beam_size 10 \
               -share_vocab \
               -dynamic_dict \
               -replace_unk \
               -model xxx.pt \
               -output xxx.txt \
               -block_ngram_repeat 3 \
               -gpu 0 \
               -min_length 400 \
               -max_length 550

Test Rouge

(1) Before ROUGE test, we should first remove special tags: .

sed -i 's/ <\/t>//g' xxx.txt
sed -i 's/<t> //g' xxx.txt

(2) Test rouge

python test_rouge.py -c summaries/xxx.txt
python test_rouge_icsi.py -c summaries/xxx.txt

xcfcode/DDAMS

DDAMS

Update

Outputs

Requirements

Data

Reproduce Results

download checkpoints

translate

remove tags

test rouge score

ROUGE score

From Scratch

Dialogue Discourse Parser

For AMI

Preprocess

Train

Translate

For ICSI

Preprocess

Train

Translate

Test Rouge