Code and data for ACL2021 paper Cross-Lingual Abstractive Summarization with Limited Parallel Resources.

Related tags

Deep Learning MCLAS
Overview

Multi-Task Framework for Cross-Lingual Abstractive Summarization (MCLAS)

The code for ACL2021 paper Cross-Lingual Abstractive Summarization with Limited Parallel Resources (Paper).

Some codes are borrowed from PreSumm (https://github.com/nlpyang/PreSumm).

Environments

Python version: This code is in Python3.7

Package Requirements: torch==1.1.0, transformers, tensorboardX, multiprocess, pyrouge

Needs few changes to be compatible with torch 1.4.0~1.8.0, mainly tensor type (bool) bugs.

Data Preparation

To improve training efficiency, we preprocessed concatenated dataset (with target "monolingual summary + [LSEP] + cross-lingual summary") and normal dataset (with target "cross-lingual summary") in advance.

You can build your own dataset or download our preprocessed dataset.

Download Preprocessed dataset.

  1. En2De dataset: Google Drive Link.
  2. En2EnDe (concatenated) dataset: Google Drive Link.
  3. En2Zh dataset: Google Drive Link.
  4. En2EnZh (concatenated) dataset: Google Drive Link.

PS: Our implementation filter some invalid samples (if the target of a sample is too short). Hence the number of the training samples may be smaller than what is reported in the paper.

Build Your Own Dataset.

Remain to be origanized. Some of the code needs to be debug, plz use it carefully.

Build tokenized files.

Plz refer to function tokenize_xgiga() or tokenize_new() in ./src/data_builder.py to write your code to preprocess your own training, validation, and test dataset. And then run the following commands:

python preprocess.py -mode tokenize_xgiga -raw_path PATH_TO_YOUR_RAW_DATA -save_path PATH_TO_YOUR_SAVE_PATH
  • Stanford CoreNLP needs to be installed.

Plz substitute tokenize_xgiga to your own process function.

In our case, we made the raw data directory as follows:

.
└── raw_directory
    ├── train
    |   ├── 1.story
    |   ├── 2.story
    |   ├── 3.story
    |   └── ...
    ├── test
    |   ├── 1.story
    |   ├── 2.story
    |   ├── 3.story
    |   └── ...
    └─ dev
        ├── 1.story
        ├── 2.story
        ├── 3.story
        └── ...

Correspondingly, the tokenized data directory is as follows

.
└── raw_directory
    ├── train
    |   ├── 1.story.json
    |   ├── 2.story.json
    |   ├── 3.story.json
    |   └── ...
    ├── test
    |   ├── 1.story.json
    |   ├── 2.story.json
    |   ├── 3.story.json
    |   └── ...
    └─ dev
        ├── 1.story.json
        ├── 2.story.json
        ├── 3.story.json
        └── ...

Build tokenized files to json files.

python preprocess.py -mode format_to_lines_new -raw_path RAW_PATH -save_path JSON_PATH -n_cpus 1 -use_bert_basic_tokenizer false -map_path MAP_PATH -shard_size 3000

Shard size is pretty important and needs to be selected carefully. This implementation use a shard as a base data unit for low-resource training. In our setting, the shard size of En2Zh, Zh2En, and En2De is 1.5k, 5k, and 3k, respectively.

Build json files to pytorch(pt) files.

python preprocess.py -mode format_to_bert_new -raw_path JSON_PATH -save_path BERT_DATA_PATH  -lower -n_cpus 1 -log_file ../logs/preprocess.log

Model Training

Full dataset scenario training

To train our model in full dataset scenario, plz use following command. Change the data path to switch the trained model between NCLS and MCLAS.

When using NCLS type datasets, arguement --multi_task enables training with NCLS+MS model.

 python train.py  \
 -task abs -mode train \
 -temp_dir ../tmp \
 -bert_data_path PATH_TO_DATA/ncls \  
 -dec_dropout 0.2  \
 -model_path ../model_abs_en2zh_noseg \
 -sep_optim true \
 -lr_bert 0.005 -lr_dec 0.2 \
 -save_checkpoint_steps 5000 \
 -batch_size 1300 \
 -train_steps 400000 \
 -report_every 50 -accum_count 5 \
 -use_bert_emb true -use_interval true \
 -warmup_steps_bert 20000 -warmup_steps_dec 10000 \
 -max_pos 512 -visible_gpus 0  -max_length 1000 -max_tgt_len 1000 \
 -log_file ../logs/abs_bert_en2zh  
 # --multi_task

Low-resource scenario training

Monolingual summarization pretraining

First we should train a monolingual summarization model using following commands:

You can change the trained model type using the same methods mentioned above (change dataset or --multi_task)

python train.py  \
-task abs -mode train \
-dec_dropout 0.2  \
-model_path ../model_abs_en2en_de/ \
-bert_data_path PATH_TO_DATA/xgiga.en \
-temp_dir ../tmp \
-sep_optim true \
-lr_bert 0.002 -lr_dec 0.2 \
-save_checkpoint_steps 2000 \
-batch_size 210 \
-train_steps 200000 \
-report_every 50 -accum_count 5 \
-use_bert_emb true -use_interval true \
-warmup_steps_bert 25000 -warmup_steps_dec 15000 \
-max_pos 512 -visible_gpus 0,1,2 -max_length 1000 -max_tgt_len 1000 \
-log_file ../logs/abs_bert_mono_enen_de \
--train_first  

# -train_from is used as continue training from certain training checkpoints.
# example:
# -train_from ../model_abs_en2en_de/model_step_70000.pt \

Low-resource scenario fine-tuning

After obtaining the monolingual model, we use it to initialize the low-resource models and continue training process.

Note:

-train_from should be omitted if you want to train a model without monolingual initialization.

--new_optim is necessary since we need to restart warm-up and learning rate decay during this process.

--few_shot controls whether to use limited resource to train the model. Meanwhile, '-few_shot_rate' controls the number of samples that you want to use. More specifically, the number of dataset's chunks.

For each scenario in our paper (using our preprocessed dataset), the few_shot_rate is set as 1, 5, and 10.

python train.py  \
-task abs -mode train \
-dec_dropout 0.2  \
-model_path ../model_abs_enende_fewshot1/ \
-train_from ../model_abs_en2en_de/model_step_50000.pt \
-bert_data_path PATH_TO_YOUR_DATA/xgiga.en \
-temp_dir ../tmp \
-sep_optim true \
-lr_bert 0.002 -lr_dec 0.2 \
-save_checkpoint_steps 1000 \
-batch_size 270 \
-train_steps 10000 \
-report_every 50 -accum_count 5 \
-use_bert_emb true -use_interval true \
-warmup_steps_bert 25000 -warmup_steps_dec 15000 \
-max_pos 512 -visible_gpus 0,2,3 -max_length 1000 -max_tgt_len 1000 \
-log_file ../logs/abs_bert_enende_fewshot1 \
--few_shot -few_shot_rate 1 --new_optim

Model Evaluation

To evaluate a model, use a command as follows:

python train.py -task abs \
-mode validate \
-batch_size 5 \
-test_batch_size 5 \
-temp_dir ../tmp \
-bert_data_path PATH_TO_YOUR_DATA/xgiga.en \
-log_file ../results/val_abs_bert_enende_fewshot1_noinit \
-model_path ../model_abs_enende_fewshot1_noinit -sep_optim true \
-use_interval true -visible_gpus 1 \
-max_pos 512 -max_length 150 \
-alpha 0.95 -min_length 20 \
-max_tgt_len 1000 \
-result_path ../logs/abs_bert_enende_fewshot1 -test_all \
--predict_2language

If you are not evaluating a MCLAS model, plz remove --predict_2language.

If you are predicting Chinese summaries, plz add --predict_chinese to the command.

If you are evaluating a NCLS+MS model, plz add --multi_task to the command.

Using following two commands will slightly improve all models' performance.

--language_limit means that the predictor will only predict words appearing in summaries of training data.

--tgt_mask is a list, recording all the words appearing in summaries of the training set. We provided chiniese and english dict in ./src directory .

Other Notable Commands

Plz ignore these arguments, these command were added and abandoned when trying new ideas¸ I will delete these related code in the future.

  • --sep_decoder
  • --few_sep_decoder
  • --tgt_seg
  • --few_sep_decoder
  • -bart

Besides, --batch_verification is used to debug, printing all the attributes in a training batch.

You might also like...
EMNLP 2021 paper Models and Datasets for Cross-Lingual Summarisation.

This repository contains data and code for our EMNLP 2021 paper Models and Datasets for Cross-Lingual Summarisation. Please contact me at [email protected]

Code for WECHSEL: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models.

WECHSEL Code for WECHSEL: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models. arXiv: https://arx

Source code and dataset for ACL2021 paper: "ERICA: Improving Entity and Relation Understanding for Pre-trained Language Models via Contrastive Learning".

ERICA Source code and dataset for ACL2021 paper: "ERICA: Improving Entity and Relation Understanding for Pre-trained Language Models via Contrastive L

Code for the ACL2021 paper
Code for the ACL2021 paper "Lexicon Enhanced Chinese Sequence Labelling Using BERT Adapter"

Lexicon Enhanced Chinese Sequence Labeling Using BERT Adapter Code and checkpoints for the ACL2021 paper "Lexicon Enhanced Chinese Sequence Labelling

Code for our paper "Sematic Representation for Dialogue Modeling" in ACL2021

AMR-Dialogue An implementation for paper "Semantic Representation for Dialogue Modeling". You may find our paper here. Requirements python 3.6 pytorch

Source code for the paper "PLOME: Pre-training with Misspelled Knowledge for Chinese Spelling Correction" in ACL2021

PLOME:Pre-training with Misspelled Knowledge for Chinese Spelling Correction (ACL2021) This repository provides the code and data of the work in ACL20

This is the code for ACL2021 paper A Unified Generative Framework for Aspect-Based Sentiment Analysis

This is the code for ACL2021 paper A Unified Generative Framework for Aspect-Based Sentiment Analysis Install the package in the requirements.txt, the

Meta Representation Transformation for Low-resource Cross-lingual Learning

MetaXL: Meta Representation Transformation for Low-resource Cross-lingual Learning This repo hosts the code for MetaXL, published at NAACL 2021. [Meta

Paddle implementation for "Cross-Lingual Word Embedding Refinement by ℓ1 Norm Optimisation" (NAACL 2021)

L1-Refinement Paddle implementation for "Cross-Lingual Word Embedding Refinement by ℓ1 Norm Optimisation" (NAACL 2021) 🙈 A more detailed readme is co

Comments
  • where is

    where is " [LSEP]" located in?

    When running the command of "Model Evaluation",my generated resulet only contains "". Could you help me find the relevant code about " [LSEP]".Thank you very much.

    opened by jiaxiansen123 3
Owner
Yu Bai
https://ybai-nlp.github.io/
Yu Bai
Code for ACL2021 paper Consistency Regularization for Cross-Lingual Fine-Tuning.

xTune Code for ACL2021 paper Consistency Regularization for Cross-Lingual Fine-Tuning. Environment DockerFile: dancingsoul/pytorch:xTune Install the f

Bo Zheng 42 Dec 9, 2022
Resources for the "Evaluating the Factual Consistency of Abstractive Text Summarization" paper

Evaluating the Factual Consistency of Abstractive Text Summarization Authors: Wojciech Kryściński, Bryan McCann, Caiming Xiong, and Richard Socher Int

Salesforce 165 Dec 21, 2022
null 190 Jan 3, 2023
Code for our paper "SimCLS: A Simple Framework for Contrastive Learning of Abstractive Summarization", ACL 2021

SimCLS Code for our paper: "SimCLS: A Simple Framework for Contrastive Learning of Abstractive Summarization", ACL 2021 1. How to Install Requirements

Yixin Liu 150 Dec 12, 2022
The code repository for EMNLP 2021 paper "Vision Guided Generative Pre-trained Language Models for Multimodal Abstractive Summarization".

Vision Guided Generative Pre-trained Language Models for Multimodal Abstractive Summarization [Paper] accepted at the EMNLP 2021: Vision Guided Genera

CAiRE 42 Jan 7, 2023
Abstractive opinion summarization system (SelSum) and the largest dataset of Amazon product summaries (AmaSum). EMNLP 2021 conference paper.

Learning Opinion Summarizers by Selecting Informative Reviews This repository contains the codebase and the dataset for the corresponding EMNLP 2021

Arthur Bražinskas 39 Jan 1, 2023
CVPR 2021 Official Pytorch Code for UC2: Universal Cross-lingual Cross-modal Vision-and-Language Pre-training

UC2 UC2: Universal Cross-lingual Cross-modal Vision-and-Language Pre-training Mingyang Zhou, Luowei Zhou, Shuohang Wang, Yu Cheng, Linjie Li, Zhou Yu,

Mingyang Zhou 28 Dec 30, 2022
Source codes for "Structure-Aware Abstractive Conversation Summarization via Discourse and Action Graphs"

Structure-Aware-BART This repo contains codes for the following paper: Jiaao Chen, Diyi Yang:Structure-Aware Abstractive Conversation Summarization vi

GT-SALT 56 Dec 8, 2022
This reporistory contains the test-dev data of the paper "xGQA: Cross-lingual Visual Question Answering".

This reporistory contains the test-dev data of the paper "xGQA: Cross-lingual Visual Question Answering".

AdapterHub 18 Dec 9, 2022
Code for the AAAI 2022 paper "Zero-Shot Cross-Lingual Machine Reading Comprehension via Inter-Sentence Dependency Graph".

multilingual-mrc-isdg Code for the AAAI 2022 paper "Zero-Shot Cross-Lingual Machine Reading Comprehension via Inter-Sentence Dependency Graph". This r

Liyan 5 Dec 7, 2022