All Projects → xlniu → Quality-Estimation1

xlniu / Quality-Estimation1

Licence: Apache-2.0 license
机器翻译子任务-翻译质量评价-复现 WMT2018 阿里论文结果

Programming Languages

python
139335 projects - #7 most used programming language
perl
6916 projects
shell
77523 projects
Jupyter Notebook
11667 projects

Projects that are alternatives of or similar to Quality-Estimation1

Quality-Estimation2
机器翻译子任务-翻译质量评价-在BERT模型后面加上Bi-LSTM进行fine-tuning
Stars: ✭ 31 (+63.16%)
Mutual labels:  nmt, quality-estimation
Nmt Keras
Neural Machine Translation with Keras
Stars: ✭ 501 (+2536.84%)
Mutual labels:  transformer, nmt
pynmt
a simple and complete pytorch implementation of neural machine translation system
Stars: ✭ 13 (-31.58%)
Mutual labels:  transformer, nmt
Njunmt Tf
An open-source neural machine translation system developed by Natural Language Processing Group, Nanjing University.
Stars: ✭ 97 (+410.53%)
Mutual labels:  transformer, nmt
Joeynmt
Minimalist NMT for educational purposes
Stars: ✭ 420 (+2110.53%)
Mutual labels:  transformer, nmt
Transformer Clinic
Understanding the Difficulty of Training Transformers
Stars: ✭ 179 (+842.11%)
Mutual labels:  transformer, nmt
sticker2
Further developed as SyntaxDot: https://github.com/tensordot/syntaxdot
Stars: ✭ 14 (-26.32%)
Mutual labels:  transformer
paccmann proteomics
PaccMann models for protein language modeling
Stars: ✭ 28 (+47.37%)
Mutual labels:  transformer
Graphormer
Graphormer is a deep learning package that allows researchers and developers to train custom models for molecule modeling tasks. It aims to accelerate the research and application in AI for molecule science, such as material design, drug discovery, etc.
Stars: ✭ 1,194 (+6184.21%)
Mutual labels:  transformer
KitanaQA
KitanaQA: Adversarial training and data augmentation for neural question-answering models
Stars: ✭ 58 (+205.26%)
Mutual labels:  transformer
dodrio
Exploring attention weights in transformer-based models with linguistic knowledge.
Stars: ✭ 233 (+1126.32%)
Mutual labels:  transformer
AdaSpeech
AdaSpeech: Adaptive Text to Speech for Custom Voice
Stars: ✭ 108 (+468.42%)
Mutual labels:  transformer
ICON
(TPAMI2022) Salient Object Detection via Integrity Learning.
Stars: ✭ 125 (+557.89%)
Mutual labels:  transformer
TDRG
Transformer-based Dual Relation Graph for Multi-label Image Recognition. ICCV 2021
Stars: ✭ 32 (+68.42%)
Mutual labels:  transformer
set-transformer
A neural network architecture for prediction on sets
Stars: ✭ 18 (-5.26%)
Mutual labels:  transformer
learningspoons
nlp lecture-notes and source code
Stars: ✭ 29 (+52.63%)
Mutual labels:  transformer
Conformer
Official code for Conformer: Local Features Coupling Global Representations for Visual Recognition
Stars: ✭ 345 (+1715.79%)
Mutual labels:  transformer
parallel-corpora-tools
Tools for filtering and cleaning parallel and monolingual corpora for machine translation and other natural language processing tasks.
Stars: ✭ 35 (+84.21%)
Mutual labels:  nmt
NLP-paper
🎨 🎨NLP 自然语言处理教程 🎨🎨 https://dataxujing.github.io/NLP-paper/
Stars: ✭ 23 (+21.05%)
Mutual labels:  transformer
Context-Transformer
Context-Transformer: Tackling Object Confusion for Few-Shot Detection, AAAI 2020
Stars: ✭ 89 (+368.42%)
Mutual labels:  transformer

Quality-Estimation1

机器翻译子任务-翻译质量评价-复现 WMT2018 阿里论文结果

简介

翻译质量评价(Quality Estimation,QE)是机器翻译领域中的一个子任务,大致可分为 Sentence-level QE,Word-level QE,Phrase-level QE,详情可参考WMT(workshop machine translation)比赛官网 http://www.statmt.org/wmt17/quality-estimation-task.html本项目针对 Sentence-level QE,试图复现论文 “Bilingual Expert” Can Find Translation Errors 的实验结果。 上述论文的开源代码如下:https://github.com/lovecambi/qebrain ,本人受服务器驱动限制,装不了高版本的tensorflow, 基于transformer开源代码 https://github.com/Kyubyong/transformer 重写了论文中的专家模型和qe模型。 由于 wmt18-qe 的测试集标签没有公布,本项目仅在 wmt17-qe 数据集上进行实验。

实验环境

python3
tensorflow == 1.2.0

实验步骤

1、准备数据
用于训练专家模型的数据,其数据来源是 WMT17 Translation task、WMT17 qe task中sentence-level task中训练集的数据(src+pe)。数据统计信息见下表:

Dataset Sentences
Europarl v7 1,920,209
Common Crawl corpus 2,399,123
News Commentary v12 268,328
Rapid corpus of EU press releases 1,314,689
UFAL Medical Corpus 2,660,314
Khresmoi development data 500
Sentence-level QE en-de smt 23,000
Sentence-level QE de-en smt 25,000
total 8,611,163
filter 8,449,133

注:过滤后的数据总量为8,449,133

2、数据预处理
tokenize(./preprocess/token.sh);
lower(./preprocess/lower.sh);
filter(./preprocess/data_filter_merge.ipynb, we filtered the source and target sentence with length <= 70 and the length ratio between 1/3 to 3);
merge(./preprocess/data_filter_merge.ipynb,将所有数据集按照语言分别合并);

3、词表生成
分别生成源端和目标端的词表,生成后的词表按照词频排序,代码见:./prepro.py,运行代码前要先设置词表大小;

4、预训练专家模型
设置exp_hyperparams.py中的参数,运行expert_model.py;

5、联合训练专家模型和qe模型
设置qe_hyperparams.py中的参数,运行qe_model.py;
使用 Sentence-level QE en-de smt 训练en-de模型,使用 Sentence-level QE de-en smt 训练de-en模型;

实验结果

Data Pearson’s vocab_size batch steps
test 2017 en-de 0.5974 30k 128 500k
state of the art(Single) 0.6837 120k - -
test 2017 de-en 0.6745 30k 128 410k
state of the art(Single) 0.7099 120k - -

注:
1、state of the art 参考论文:“Bilingual Expert” Can Find Translation Errors
2、表中的batch和steps都是说的训练专家模型的情况,steps指训练了多少步(每一步是一个batch);
3、从我的观察来看,基本上是专家模型被训练的steps越多,后续再联合训练qe模型而得到的结果就越好,但是随着专家模型被训练步数的不断增多(比如达到350k以上),最终的结果会出现波动;
4、训练专家模型和联合训练qe模型的日志文件都放在了 ./nohup 文件夹下;

Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].