information-extraction

2019 Language and Intelligence Challenge: Information Extraction

Prerequisites

• Install required packages by:

pip install -r requirements.txt

Data

Download data: initialize and update the information-extraction-data git submodule by git submodule init and git submodule update, and then unzip the data files

• sample schema:

  {"object_type": "地点", "predicate": "祖籍", "subject_type": "人物"}
  

• sample data, with postag and text as input and spo_list as output:

  {
      "postag": [
          {"word": "一直", "pos": "d"}, 
          {"word": "陪", "pos": "v"}, 
          {"word": "我", "pos": "r"}, 
          {"word": "到", "pos": "p"}, 
          {"word": "现在", "pos": "t"}, 
          {"word": "是", "pos": "v"}, 
          {"word": "歌手", "pos": "n"}, 
          {"word": "马健涛", "pos": "nr"}, 
          {"word": "原创", "pos": "v"}, 
          {"word": "的", "pos": "u"}, 
          {"word": "歌曲", "pos": "n"}
      ], 
      "text": "一直陪我到现在是歌手马健涛原创的歌曲", 
      "spo_list": [
          {"predicate": "歌手", "object_type": "人物", "subject_type": "歌曲", "object": "马健涛", "subject": "一直陪我到现在"}
      ]
  }
  

Baseline

• baidu/information-extraction

Idea

• Train multi-label classification model: predict predicate.
• Train sequence labeling model: input text and predicate, output text labeling.
• Extract SPO from sequence labeling result.

Implementation

Check report/PRML-final-project-doc-2019.pdf for details.

Multi-label Classification

• CNN, BiRNN, BiLSTM, BiLSTM with max pooling and RCNN
• BERT

Sequence Labeling

• Encoder: BiLSTM and Transformer
• Decoder: CRF

Result

Multi-label Classification

Sequence Labeling

fitlog usage

• Initialize fitlog in classification folder:

cd classification/
fitlog init
fitlog log logs

• Initialize fitlog in labeling folder:

cd labeling/
fitlog init
fitlog log logs

Author

Zhongyu Chen