Skip to content

Latest commit

 

History

History
130 lines (103 loc) · 6.77 KB

README.md

File metadata and controls

130 lines (103 loc) · 6.77 KB
Raw

ime

Transformers4IME

*其他语言版本: English

Transformers4IME是尝试将预训练模型运用于输入法的软件包。

PinyinGPT

PinyinGPT模型源于我们发表于ACL2022的工作 Exploring and Adapting Chinese GPT to Pinyin Input Method 。

@inproceedings{tan-etal-2022-exploring,
    title = "Exploring and Adapting {C}hinese {GPT} to {P}inyin Input Method",
    author = "Tan, Minghuan  and
      Dai, Yong  and
      Tang, Duyu  and
      Feng, Zhangyin  and
      Huang, Guoping  and
      Jiang, Jing  and
      Li, Jiwei  and
      Shi, Shuming",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.acl-long.133",
    doi = "10.18653/v1/2022.acl-long.133",
    pages = "1899--1909",
    abstract = "While GPT has become the de-facto method for text generation tasks, its application to pinyin input method remains unexplored.In this work, we make the first exploration to leverage Chinese GPT for pinyin input method.We find that a frozen GPT achieves state-of-the-art performance on perfect pinyin.However, the performance drops dramatically when the input includes abbreviated pinyin.A reason is that an abbreviated pinyin can be mapped to many perfect pinyin, which links to even larger number of Chinese characters.We mitigate this issue with two strategies,including enriching the context with pinyin and optimizing the training process to help distinguish homophones. To further facilitate the evaluation of pinyin input method, we create a dataset consisting of 270K instances from fifteen domains.Results show that our approach improves the performance on abbreviated pinyin across all domains.Model analysis demonstrates that both strategiescontribute to the performance boost.",
}

本文主要研究了将中文GPT的预训练模型适配到拼音输入法的问题。我们发现，在GPT的广泛使用中，仍然缺少对拼音输入法的探索。 经过对生成过程加上拼音的限制，全拼场景下的GPT的效果十分突出，在传统的数据集上就能达到SOTA。 然而，对于首字母的情形，GPT的效果出现大幅下滑，这与同声母字的候选大幅增加相关。 我们采取两种策略来解决这个问题，一方面让模型充分使用上下文信息和拼音信息，另一方面增强训练过程中对同声母字的辨析。 为了助力拼音输入法的评测，我们基于最新的语料，构建了跨15个新闻领域的270k的测试集合，集合的样本覆盖多种上文的长度和预测长度组合。 通过对模型的分析和消融，我们发现模型的两个策略都对最后的效果有促进作用。 实验结果对输入法的研究具有参考意义。

pinyinGPT-method

语料整理

例如，处理拼音的相关语料时, 我们会得到如下数据格式

{'words': [['观众', '姥爷'], ['，'], ['如果', '你', '有', '超神', '超', '秀'], ['、'], ['坑爹', '搞笑', '素材'], ['，'],
           ['欢迎', '给', '苍', '姐', '投稿'], ['，'], ['采用', '有奖', '哦'], ['！']],
 'tokens': [[['观', '众'], ['姥', '爷']], ['，'], [['如', '果'], ['你'], ['有'], ['超', '神'], ['超'], ['秀']], ['、'],
            [['坑', '爹'], ['搞', '笑'], ['素', '材']], ['，'], [['欢', '迎'], ['给'], ['苍'], ['姐'], ['投', '稿']], ['，'],
            [['采', '用'], ['有', '奖'], ['哦']], ['！']],
 'pinyin': [[['guan', 'zhong'], ['lao', 'ye']], ['，'],
            [['ru', 'guo'], ['ni'], ['you'], ['chao', 'shen'],
             ['chao'], ['xiu']], ['、'],
            [['keng', 'die'], ['gao', 'xiao'], ['su', 'cai']],
            ['，'], [['huan', 'ying'], ['gei'], ['cang'], ['jie'],
                    ['tou', 'gao']], ['，'],
            [['cai', 'yong'], ['you', 'jiang'], ['o']], ['！']],
 'abbr': [[['g', 'z'], ['l', 'y']], ['，'], [['r', 'g'], ['n'], ['y'], ['c', 's'], ['c'], ['x']], ['、'],
          [['k', 'd'], ['g', 'x'], ['s', 'c']], ['，'], [['h', 'y'], ['g'], ['c'], ['j'], ['t', 'g']], ['，'],
          [['c', 'y'], ['y', 'j'], ['o']], ['！']]}

选定需要的db文件进行合并通过transformers支持的tokenizer转换成token id，得到一个模型可直接使用的txt_db。

PYTHONPATH=src RAW_DIR=data/raw ANNOTATION_DIR=data/annotations_db TXT_DIR=data/txt_db python convert.py --domain CLUECorpusSmall --genre news2016zh_corpus --config=config/gpt2zh/pretrain_pinyin.json --use_proxy
PYTHONPATH=src RAW_DIR=data/raw ANNOTATION_DIR=data/annotations_db2 TXT_DIR=data/txt_db ANNOTATOR_TAGGER=whitespace ADDITIONAL_SPECIAL_TOKENS=data/pretrained/additional_special_tokens.json PRETRAINED_MODEL_NAME_OR_PATH=data/pretrained/uer/gpt2-chinese-cluecorpussmall python convert.py --domain 300g_word --genre train.txt07 --config=config/gpt2zh/pretrain_pinyin.json --use_proxy --split train

模型列表

  • GPT2
  • PinyinGPT2Concat
    • Directly
    • Segmented (visualjoyce/transformers4ime-pinyingpt-concat) 🤗 models
  • PinyinGPT2Compose
    • PinyinGPT2ComposeBottom
    • PinyinGPT2ComposeTop
      • logits
      • states
      • residual

训练模式

  • AbbrOnly 全缩写
  • PinyinOnly 全拼音
  • PinyinAbbr 混合模式
sh pretrain_pinyingpt.sh

基线测试

基线评测数据集地址

99E333F0B1C6D7B67ACB9D9E61A73DA8

PD基线测试

python3 benchmarks.py --samples_json data/benchmarks/PD/samples_0.json \
  --pretrained_model_name_or_path data/pretrained_models/gpt2-zh-ours \
  --additional_special_tokens data/pretrained/additional_special_tokens.json \
  --pinyin2char_json data/pretrained/pinyin2char.json \
  --pinyin_logits_processor_cls pinyingpt-compatible \
  --num_beams 16 \
  --abbr_mode none

支持对特定模型的特定checkpoint进行评测

sh benchmarks.sh pinyingpt-concat data/output/pinyingpt \
  data/output/models/ckpt50000/pytorch_model.bin

鸣谢

该工作在腾讯AI Lab实习期间完成。