Code for our paper "Transfer Learning for Sequence Generation: from Single-source to Multi-source" in ACL 2021.

Overview

TRICE: a task-agnostic transferring framework for multi-source sequence generation

This is the source code of our work Transfer Learning for Sequence Generation: from Single-source to Multi-source (ACL 2021).

We propose TRICE, a task-agnostic Transferring fRamework for multI-sourCe sEquence generation, for transferring pretrained models to multi-source sequence generation tasks (e.g., automatic post-editing, multi-source translation, and multi-document summarization). TRICE achieves new state-of-the-art results on the WMT17 APE task and the multi-source translation task using the WMT14 test set. Welcome to take a quick glance at our blog.

The implementation is on top of the open-source NMT toolkit THUMT.

@misc{huang2021transfer,
      title={Transfer Learning for Sequence Generation: from Single-source to Multi-source}, 
      author={Xuancheng Huang and Jingfang Xu and Maosong Sun and Yang Liu},
      year={2021},
      eprint={2105.14809},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Contents

Prerequisites

  • Python >= 3.6
  • tensorflow-cpu >= 2.0
  • torch >= 1.7
  • transformers >= 3.4
  • sentencepiece >= 0.1

Pretrained model

We adopt mbart-large-cc25 in our experiments. Other sequence-to-sequence pretrained models can also be used with only a few modifications.

If your GPUs do not have enough memories, you can prune the original large vocabulary (25k) to a small vocabulary (e.g., 3k) with little performance loss.

Finetuning

Single-source finetuning

PYTHONPATH=${path_to_TRICE} \
python ${path_to_TRICE}/thumt/bin/trainer.py \
    --input ${train_src1} ${train_src2} ${train_trg} \
    --vocabulary ${vocab_joint} ${vocab_joint} \
    --validation ${dev_src1} ${dev_src2} \
    --references ${dev_ref} \
    --model transformer --half --hparam_set big \
    --output single_finetuned \
    --parameters \
fixed_batch_size=false,batch_size=820,train_steps=120000,update_cycle=5,device_list=[0,1,2,3],\
keep_checkpoint_max=2,save_checkpoint_steps=2000,\
eval_steps=2001,decode_alpha=1.0,decode_batch_size=16,keep_top_checkpoint_max=1,\
attention_dropout=0.1,relu_dropout=0.1,residual_dropout=0.1,learning_rate=5e-05,warmup_steps=4000,initial_learning_rate=5e-8,\
separate_encode=false,separate_cross_att=false,segment_embedding=false,\
input_type="single_random",adapter_type="None",num_fine_encoder_layers=0,normalization="after",\
src_lang_tok="en_XX",hyp_lang_tok="de_DE",tgt_lang_tok="de_DE",mbart_model_code="facebook/mbart-large-cc25",\
spm_path="sentence.bpe.model",pad="<pad>",bos="<s>",eos="</s>",unk="<unk>"

Multi-source finetuning

PYTHONPATH=${path_to_TRICE} \
python ${path_to_TRICE}/thumt/bin/trainer.py \
    --input ${train_src1} ${train_src2} ${train_tgt} \
    --vocabulary ${vocab_joint} ${vocab_joint} \
    --validation ${dev_src1} ${dev_src2} \
    --references ${dev_ref} \
    --model transformer --half --hparam_set big \
    --checkpoint single_finetuned/eval/model-best.pt \
    --output multi_finetuned \
    --parameters \
fixed_batch_size=false,batch_size=820,train_steps=120000,update_cycle=5,device_list=[0,1,2,3],\
keep_checkpoint_max=2,save_checkpoint_steps=2000,\
eval_steps=2001,decode_alpha=1.0,decode_batch_size=16,keep_top_checkpoint_max=1,\
attention_dropout=0.1,relu_dropout=0.1,residual_dropout=0.1,learning_rate=5e-05,warmup_steps=4000,initial_learning_rate=5e-8,special_learning_rate=5e-04,special_var_name="adapter",\
separate_encode=false,separate_cross_att=true,segment_embedding=true,\
input_type="",adapter_type="Cross-attn",num_fine_encoder_layers=1,normalization="after",\
src_lang_tok="en_XX",hyp_lang_tok="de_DE",tgt_lang_tok="de_DE",mbart_model_code="facebook/mbart-large-cc25",\
spm_path="sentence.bpe.model",pad="<pad>",bos="<s>",eos="</s>",unk="<unk>"

Arguments to be explained

** special_learning_rate: if a variable's name contains special_var_name, the learning rate of it will be special_learning_rate. We give the fine encoder a larger learning rate.
** separate_encode: whether to encode multiple sources separately before the fine encoder.
** separate_cross_att: whether to use separated cross-attention described in our paper.
** segment_embedding: whether to use sinusoidal segment embedding described in our paper.
** input_type: "single_random" for single-source finetuning , "" for multi-source finetuning.
** adapter_type: "None" for no fine encoder, "Cross-attn" for fine encoder with cross-attention.
** num_fine_encoder_layers: number of fine encoder layers.
** src_lang_tok: language token for the first source sentence. Please refer to here for language tokens for all 25 languages.
** hyp_lang_tok: language token for the second source sentence.
** tgt_lang_tok: language token for the target sentence.
** mbart_model_code: model code for transformers.
** spm_path: sentence piece model (can download from here).

Explanations for other arguments could be found in the user manual of THUMT.

Inference

PYTHONPATH=${path_to_TRICE} \
python ${path_to_TRICE}/thumt/bin/translator.py \
  --input ${test_src1} ${test_src2} --output ${test_tgt} \
  --vocabulary ${vocab_joint} ${vocab_joint} \
  --checkpoints multi_finetuned/eval/model-best.pt \
  --model transformer --half \
  --parameters device_list=[0,1,2,3],decode_alpha=1.0,decode_batch_size=32
# recover sentence piece tokenization
...
# calculate BLEU
...

Contact

If you have questions, suggestions and bug reports, please email [email protected].

Owner
THUNLP-MT
Machine Translation Group, Natural Language Processing Lab at Tsinghua University (THUNLP). Please refer to https://github.com/thunlp for more NLP resources.
THUNLP-MT
This repo contains simple to use, pretrained/training-less models for speaker diarization.

PyDiar This repo contains simple to use, pretrained/training-less models for speaker diarization. Supported Models Binary Key Speaker Modeling Based o

12 Jan 20, 2022
jiant is an NLP toolkit

🚨 Update 🚨 : As of 2021/10/17, the jiant project is no longer being actively maintained. This means there will be no plans to add new models, tasks,

ML² AT CILVR 1.5k Dec 28, 2022
🚀 RocketQA, dense retrieval for information retrieval and question answering, including both Chinese and English state-of-the-art models.

In recent years, the dense retrievers based on pre-trained language models have achieved remarkable progress. To facilitate more developers using cutt

475 Jan 04, 2023
Code for EMNLP'21 paper "Types of Out-of-Distribution Texts and How to Detect Them"

Code for EMNLP'21 paper "Types of Out-of-Distribution Texts and How to Detect Them"

Udit Arora 19 Oct 28, 2022
PyTorch implementation of "data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language" from Meta AI

data2vec-pytorch PyTorch implementation of "data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language" from Meta AI (F

Aryan Shekarlaban 105 Jan 04, 2023
The proliferation of disinformation across social media has led the application of deep learning techniques to detect fake news.

Fake News Detection Overview The proliferation of disinformation across social media has led the application of deep learning techniques to detect fak

Kushal Shingote 1 Feb 08, 2022
Athena is an open-source implementation of end-to-end speech processing engine.

Athena is an open-source implementation of end-to-end speech processing engine. Our vision is to empower both industrial application and academic research on end-to-end models for speech processing.

Ke Technologies 34 Sep 08, 2022
DANeS is an open-source E-newspaper dataset by collaboration between DATASET JSC (dataset.vn) and AIV Group (aivgroup.vn)

DANeS - Open-source E-newspaper dataset Source: Technology vector created by macrovector - www.freepik.com. DANeS is an open-source E-newspaper datase

DATASET .JSC 64 Aug 17, 2022
ACL22 paper: Imputing Out-of-Vocabulary Embeddings with LOVE Makes Language Models Robust with Little Cost

Imputing Out-of-Vocabulary Embeddings with LOVE Makes Language Models Robust with Little Cost LOVE is accpeted by ACL22 main conference as a long pape

Lihu Chen 32 Jan 03, 2023
Non-Autoregressive Translation with Layer-Wise Prediction and Deep Supervision

Deeply Supervised, Layer-wise Prediction-aware (DSLP) Transformer for Non-autoregressive Neural Machine Translation

Chenyang Huang 37 Jan 04, 2023
Yet Another Compiler Visualizer

yacv: Yet Another Compiler Visualizer yacv is a tool for visualizing various aspects of typical LL(1) and LR parsers. Check out demo on YouTube to see

Ashutosh Sathe 129 Dec 17, 2022
Russian words synonyms and antonyms

ru_synonyms Russian words synonyms and antonyms. Install pip install git+https://github.com/ahmados/rusynonyms.git Usage from ru_synonyms import Anto

sumekenov 7 Dec 14, 2022
Implementing SimCSE(paper, official repository) using TensorFlow 2 and KR-BERT.

KR-BERT-SimCSE Implementing SimCSE(paper, official repository) using TensorFlow 2 and KR-BERT. Training Unsupervised python train_unsupervised.py --mi

Jeong Ukjae 27 Dec 12, 2022
This project uses unsupervised machine learning to identify correlations between daily inoculation rates in the USA and twitter sentiment in regards to COVID-19.

Twitter COVID-19 Sentiment Analysis Members: Christopher Bach | Khalid Hamid Fallous | Jay Hirpara | Jing Tang | Graham Thomas | David Wetherhold Pro

4 Oct 15, 2022
Unifying Cross-Lingual Semantic Role Labeling with Heterogeneous Linguistic Resources (NAACL-2021).

Unifying Cross-Lingual Semantic Role Labeling with Heterogeneous Linguistic Resources Description This is the repository for the paper Unifying Cross-

Sapienza NLP group 16 Sep 09, 2022
GraphNLI: A Graph-based Natural Language Inference Model for Polarity Prediction in Online Debates

GraphNLI: A Graph-based Natural Language Inference Model for Polarity Prediction in Online Debates Vibhor Agarwal, Sagar Joglekar, Anthony P. Young an

Vibhor Agarwal 2 Jun 30, 2022
Simple and efficient RevNet-Library with DeepSpeed support

RevLib Simple and efficient RevNet-Library with DeepSpeed support Features Half the constant memory usage and faster than RevNet libraries Less memory

Lucas Nestler 112 Dec 05, 2022
English loanwords in the world's languages

Wiktionary as CLDF Content cldf1 and cldf2 contain cldf-conform data sets with a total of 2 377 756 entries about the vocabulary of all 1403 languages

Viktor Martinović 3 Jan 14, 2022
Espial is an engine for automated organization and discovery of personal knowledge

Live Demo (currently not running, on it) Espial is an engine for automated organization and discovery in knowledge bases. It can be adapted to run wit

Uzay-G 159 Dec 30, 2022
AI and Machine Learning workflows on Anthos Bare Metal.

Hybrid and Sovereign AI on Anthos Bare Metal Table of Contents Overview Terraform as IaC Substrate ABM Cluster on GCE using Terraform TensorFlow ResNe

Google Cloud Platform 8 Nov 26, 2022