Git Product home page Git Product logo

deepie's Introduction

DeepIE: Deep Learning for Information Extraction

DeepIE: 基于深度学习的信息抽取技术(预计2020年8月31日前全部更新完毕)

TOP

Papers

Codes

1. 实体抽取

  • 各主流方法在主要中文NER数据集上的表现情况 具体说明
lexicon Ontonotes MSRA Resume Weibo
biLSTM ---- 71.81 91.87 94.41 56.75
Lattice LSTM 词表1 73.88 93.18 94.46 58.79
WC-LSTM 词表1 74.43 93.36 94.96 49.86
LR-CNN 词表1 74.45 93.71 95.11 59.92
CGN 词表2 74.79 93.47 94.12 63.09
LGN 词表1 74.85 93.63 95.41 60.15
Simple-Lexicon 词表1 75.54 93.50 95.59 61.24
FLAT 词表1 76.45 94.12 95.45 60.32
FLAT 词表2 75.70 94.35 94.93 63.42
BERT ---- 80.14 94.95 95.53 68.20
BERT+FLAT 词表1 81.82 96.09 95.86 68.55
  • MSRA-NER
方法 f p r
char+ lstm-crf 86.18% 88.43% 83.10%
char-bigram + lstm-crf 91.80% 92.60% 90.34%
char-bigram + adTransformer-crf 92.98% 93.25% 92.72%
char-bigram + lexion-augment + lstm-crf 93.33% 94.26% 92.43%
char-bigram-BERT + lstm-crf 94.71% 95.14% 94.27%
char-bigram-BERT + lexion-augment + lstm-crf 95.26% 95.90% 94.63%
  • CCKS2019-医疗实体抽取
方法 f p r
char-bigram + lstm-crf 81.76% 82.91% 80.6
+ domain transfer(from ccks2018 to 2019) 82.54% 83.43% 81.81%
char-bigram + adTransformer-crf 82.83% 82.19% 83.49%
char-bigram + lexion-augment + lstm-crf 82.76% 82.79% 82.72%
BERT-finetune+crf 83.49% 84.11% 82.89%
roBERTa-finetune+crf 83.66% 83.67% 83.66%
char-bigram-BERT + lstm-crf 83.37% 83.51% 83.22%
char-bigram-BERT + lexion-augment + lstm-crf 84.15% 84.29% 84.01%
  • CCKS2020-医疗实体抽取

(注:测试集与ccks2019一致,去除ccks2020训练集中已经在2019测试集中的样本,下列指标未做规则处理和模型融合)

方法 f p r
char-bigram + lstm-crf 82.68% 83.14% 82.22%
char-bigram + lexion-augment + lstm-crf 83.12% 83.10% 83.14%
char-bigram-BERT + lstm-crf 83.12% 83.04% 83.21%
char-bigram-BERT-RoBerta_wwm + lstm-crf 83.66% 83.76% 83.56%
char-bigram-BERT-XLNet + lstm-crf 84.12% 83.88% 84.36%
char-bigram-BERT + lexion-augment + lstm-crf 84.50% 84.32% 84.67%
  • CCKS2020-面向试验鉴定的命名实体识别任务:TODO

2. 实体关系联合抽取

具体使用说明

  • 2019语言与智能技术竞赛:关系抽取任务
方法 f(dev) p(dev) r(dev)
multi head selection 76.36 79.24 73.69
ETL-BIES 77.07% 77.13% 77.06%
ETL-Span 78.94% 80.11% 77.8%
ETL-Span + word2vec 79.99% 80.62% 79.38%
ETL-Span + word2vec + adversarial training 80.38% 79.95% 80.82%
ETL-Span + BERT 81.88% 82.35% 81.42%
  • 2020语言与智能技术竞赛:关系抽取任务
方法 f(dev) p(dev) r(dev)
ETL-Span + BERT 74.58 74.44 74.71

3. 属性抽取

  • 领域数据集:瑞金医院糖尿病信息抽取数据
# 药物-属性
['药品-用药频率','药品-持续时间','药品-用药剂量','药品-用药方法','药品-不良反应']
# 疾病-属性
['疾病-检查方法','疾病-临床表现','疾病-非药治疗','疾病-药品名称','疾病-部位']
主体 方法 f p r
疾病 lstm+ multi-label pointer network 76.55 74.36 78.86
疾病 bert + multi-label pointer network 77.59 77.45 77.74
药物 lstm+ multi-label pointer network 81.12 79.15 83.19

4. 实体链接/标准化

5.事件抽取

  • CCKS2020-医疗事件抽取

  • CCKS2020:面向金融领域的篇章级事件主体抽取

  • CCKS2020:面向金融领域的篇章级事件要素抽取

6.信息抽取中的低资源解决方案

TODO-list

  • 信息抽取领域的数据资源汇总:

    • 医疗
    • 金融
    • 电商
    • 法律
  • 信息抽取相关竞赛汇总:

    • 百度-2020语言与智能技术竞赛:关系抽取任务
    • 百度-2020语言与智能技术竞赛:事件抽取任务
    • 百度-2019语言与智能技术竞赛:信息抽取
    • CCKS 2019 医疗命名实体识别
    • CHIP 2019 临床术语标准化任务
    • CCKS 2019 人物关系抽取
    • CCKS 2019 公众公司公告信息抽取
    • CCKS 2019 面向金融领域的事件主体抽取
  • 摘要抽取

  • 前沿技术在信息抽取中的应用

Reference

deepie's People

Contributors

loujie0822 avatar

Watchers

 avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.