A Baseline System For CCKS-2019-IPRE

Introduction

竞赛官网：地址 当前模型：CNN with selective attention. （由官方提供地址）

Getting Started

Environment Requirements

• python 3.6
• numpy
• tensorflow 1.12.0

Step 1: 下载数据

Please download the data from the competition website, then unzip files and put them in ./data/ folder.

Step 2: 数据清洗及训练词向量

python pretrain_embedding.py

• raw_file： 文本语料
• clean_file： 清洗后语料
• seg_file： 分词后语料
• word2vec.txt：词向量文件

Step 3: 生成词典和词向量矩阵的pickle文件

python baseline.py --mode prepare

Then you will get two pickle files in './data/processed'

• word_map.pkl
• word_embed.pkl

Step 4: 训练模型

You can use the following command to train models for Sent-Track or Bag-Track:

python baseline.py --level sent 
python baseline.py --level bag

The model will be stored in ./model/ floder. We provide large scale unmarked corpus for train word vectors or language mdoels. The word vectors used in baseline system are trained by a package named gensim in python, and some parameters are set as follows:

from gensim.models import word2vec
model = word2vec.Word2Vec(sentences, sg=1, size=300, window=5, min_count=10, negative=5, sample=1e-4, workers=10)

Step 5: 测试模型

You can use the following command to test models for Sent-Track or Bag-Track:

python baseline.py --mode test --level sent 
python baseline.py --mode test --level bag

Predicted results will be stored in result_sent.txt or result_bag.txt.

结果

我们的设备Intel 5118(12核) + 64GB RAM + RTX 2080Ti

模型	线下sent	线上sent	线下bag	线上bag	训练时长
官方（报告）	-	0.22	-	0.31	4h40min(sent)
CNN with attention	0.226087	0.21564			2h45min(sent)

参考文献

• Lin Y, Shen S, Liu Z, et al. Neural relation extraction with selective attention over instances[C]//Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2016, 1: 2124-2133.