Official implementation of Exploiting Entity BIO Tag Embeddings and Multi-task Learning for Relation Extraction with Imbalanced Data.

Contributed by National Engineering Research Center for Software Engineering, Peking University.

Overview

This is a TensorFlow-based framework for Relation Extraction using BIO tag embeddings and multi-task learning, we use Keras to easily implement our methods. Our model has three parts:

• Input Layer:
  • Word Embeddings
  • Positional Embeddings
  • BIO tag Embeddings
• Convolutional Layer with Multi-Sized Window Kernels
• Multi-Task Layer:
  • Relation Identification with Cross-entropy Loss
  • Relation Classification with Ranking Loss

Data Generation and Parameter Settings

The dataset we used in this paper is AEC2005 (English corpus and Chinese corpus), which is a very popular dataset for relation extraction. The Data Preparation and Parameter Settings are mentioned in our paper, we will also release the processed data later to facilitate future research.

Requirements

• Python(3.6)
• Numpy(>=1.13.3)
• Tensorflow (>=1.9)
• Keras(>=2.1.1)
• scikit-learn(>=0.18)

Test Results

English Corpus:

Model	P%	R%	F1%
SPTree	70.1	61.2	65.3
Walk-based	69.7	59.5	64.2
Baseline	58.8	57.3	57.2
Baseline+Tag	61.3	76.7	67.4
Baseline+MTL	63.8	56.1	59.5
Baseline+MTL+Tag	66.5	71.8	68.9

Chinese Corpus:

Model	P%	R%	F1%
PCNN	54.4	42.1	46.1
Eatt-BiGRU	57.8	49.7	52.0
Baseline	48.5	57.1	51.7
Baseline+Tag	61.8	62.7	61.4
Baseline+MTL	56.7	52.9	53.8
Baseline+MTL+Tag	61.3	65.8	62.9

Citation

@inproceedings{ye-etal-2019-exploiting,
    title = "Exploiting Entity {BIO} Tag Embeddings and Multi-task Learning for Relation Extraction with Imbalanced Data",
    author = "Ye, Wei  and
      Li, Bo  and
      Xie, Rui  and
      Sheng, Zhonghao  and
      Chen, Long  and
      Zhang, Shikun",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P19-1130",
    doi = "10.18653/v1/P19-1130",
    pages = "1351--1360"
}