Codes for our IJCAI21 paper: Dialogue Discourse-Aware Graph Model and Data Augmentation for Meeting Summarization

Related tags

Deep LearningDDAMS
Overview

DDAMS

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

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.

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

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
Owner
xcfeng
Ph.D. candidate working on Summarization.
xcfeng
GitHub Repository
We present a framework for training multi-modal deep learning models on unlabelled video data by forcing the network to learn invariances to transformations applied to both the audio and video streams.

Multi-Modal Self-Supervision using GDT and StiCa This is an official pytorch implementation of papers: Multi-modal Self-Supervision from Generalized D

Facebook Research 42 Dec 09, 2022
PyTorch implementation of Advantage Actor Critic (A2C), Proximal Policy Optimization (PPO), Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation (ACKTR) and Generative Adversarial Imitation Learning (GAIL).

pytorch-a2c-ppo-acktr Update (April 12th, 2021) PPO is great, but Soft Actor Critic can be better for many continuous control tasks. Please check out

Ilya Kostrikov 3k Jan 09, 2023
Code for generating a single image pretraining dataset

Single Image Pretraining of Visual Representations As shown in the paper A critical analysis of self-supervision, or what we can learn from a single i

Yuki M. Asano 12 Dec 19, 2022
Official code for "Distributed Deep Learning in Open Collaborations" (NeurIPS 2021)

Distributed Deep Learning in Open Collaborations This repository contains the code for the NeurIPS 2021 paper "Distributed Deep Learning in Open Colla

Yandex Research 96 Sep 15, 2022
Official implementation of "UCTransNet: Rethinking the Skip Connections in U-Net from a Channel-wise Perspective with Transformer"

[AAAI2022] UCTransNet This repo is the official implementation of "UCTransNet: Rethinking the Skip Connections in U-Net from a Channel-wise Perspectiv

Haonan Wang 199 Jan 03, 2023
RSNA Intracranial Hemorrhage Detection with python

RSNA Intracranial Hemorrhage Detection This is the source code for the first place solution to the RSNA2019 Intracranial Hemorrhage Detection Challeng

24 Nov 30, 2022
Python scripts to detect faces in Python with the BlazeFace Tensorflow Lite models

Python scripts to detect faces using Python with the BlazeFace Tensorflow Lite models. Tested on Windows 10, Tensorflow 2.4.0 (Python 3.8).

Ibai Gorordo 46 Nov 17, 2022
Fast, modular reference implementation and easy training of Semantic Segmentation algorithms in PyTorch.

TorchSeg This project aims at providing a fast, modular reference implementation for semantic segmentation models using PyTorch. Highlights Modular De

ycszen 1.4k Jan 02, 2023
Real-time VIBE: Frame by Frame Inference of VIBE (Video Inference for Human Body Pose and Shape Estimation)

Real-time VIBE Inference VIBE frame-by-frame. Overview This is a frame-by-frame inference fork of VIBE at [https://github.com/mkocabas/VIBE]. Usage: i

23 Jul 02, 2022
This repository is based on Ultralytics/yolov5, with adjustments to enable polygon prediction boxes.

Polygon-Yolov5 This repository is based on Ultralytics/yolov5, with adjustments to enable polygon prediction boxes. Section I. Description The codes a

xinzelee 226 Jan 05, 2023
[CVPR 2021] NormalFusion: Real-Time Acquisition of Surface Normals for High-Resolution RGB-D Scanning

NormalFusion: Real-Time Acquisition of Surface Normals for High-Resolution RGB-D Scanning Project Page | Paper | Supplemental material #1 | Supplement

KAIST VCLAB 49 Nov 24, 2022
🚩🚩🚩

My CTF Challenges 2021 AIS3 Pre-exam / MyFirstCTF Name Category Keywords Difficulty ⒸⓄⓋⒾⒹ-①⑨ (MyFirstCTF Only) Reverse Baby ★ Piano Reverse C#, .NET ★

6 Oct 28, 2021
Inkscape extensions for figure resizing and editing

Academic-Inkscape: Extensions for figure resizing and editing This repository contains several Inkscape extensions designed for editing plots. Scale P

192 Dec 26, 2022
Multi-Task Deep Neural Networks for Natural Language Understanding

New Release We released Adversarial training for both LM pre-training/finetuning and f-divergence. Large-scale Adversarial training for LMs: ALUM code

Xiaodong 2.1k Dec 30, 2022
Data Consistency for Magnetic Resonance Imaging

Data Consistency for Magnetic Resonance Imaging Data Consistency (DC) is crucial for generalization in multi-modal MRI data and robustness in detectin

Dimitris Karkalousos 19 Dec 12, 2022
PyTorch implementation of Trust Region Policy Optimization

PyTorch implementation of TRPO Try my implementation of PPO (aka newer better variant of TRPO), unless you need to you TRPO for some specific reasons.

Ilya Kostrikov 366 Nov 15, 2022
Pytorch implementation of Implicit Behavior Cloning.

Implicit Behavior Cloning - PyTorch (wip) Pytorch implementation of Implicit Behavior Cloning. Install conda create -n ibc python=3.8 pip install -r r

Kevin Zakka 49 Dec 25, 2022
Sound Source Localization for AI Grand Challenge 2021

Sound-Source-Localization Sound Source Localization study for AI Grand Challenge 2021 (sponsored by NC Soft Vision Lab) Preparation 1. Place the data-

sanghoon 19 Mar 29, 2022
(CVPR 2022) Pytorch implementation of "Self-supervised transformers for unsupervised object discovery using normalized cut"

(CVPR 2022) TokenCut Pytorch implementation of Tokencut: Self-supervised Transformers for Unsupervised Object Discovery using Normalized Cut Yangtao W

YANGTAO WANG 200 Jan 02, 2023
Neural Architecture Search Powered by Swarm Intelligence 🐜

Neural Architecture Search Powered by Swarm Intelligence 🐜 DeepSwarm DeepSwarm is an open-source library which uses Ant Colony Optimization to tackle

288 Oct 28, 2022