融合多层次特征的中文语义角色标注

PDF下载HTML

Title:

Chinese semantic role labeling with multi-level linguistic features

作者:

王一成^1,2, 万福成¹, 马宁²

1. 西北民族大学 中国民族语言文字信息技术教育部重点实验室, 甘肃 兰州 730030;
2. 西北民族大学 甘肃省民族语言智能处理重点实验室, 甘肃 兰州 730030

Author(s):

WANG Yicheng^1,2, WAN Fucheng¹, MA Ning²

1. Key Laboratory of China’s Ethnic Languages and Information Technology of Ministry of Education, Northwest Minzu University, Lanzhou, Gansu 730030, China;
2. Key Laboratory of China’s Ethnic Languages and Intelligent Processing of Gansu Province, Nor

关键词:

自然语言处理; 语义角色标注; 深度学习; Bi-LSTM; 语言学特征; 后处理层; Max pooling

Keywords:

natural language processing; semantic role labeling; deep learning; Bi-LSTM; linguistic characteristics; post-processing layer; Max pooling

分类号:

TP391

DOI:

10.11992/tis.201910012

摘要:

随着人工智能和中文信息处理技术的迅猛发展，自然语言处理相关研究已逐步深入到语义理解层次上，而中文语义角色标注则是语义理解领域的核心技术。在统计机器学习仍占主流的中文信息处理领域，传统的标注方法对句子的句法及语义的解析程度依赖较大，因而标注准确率受限较大，已无法满足当前需求。针对上述问题，对基于Bi-LSTM的中文语义角色标注基础模型进行了改进研究，在模型后处理阶段结合了Max pooling技术，训练时融入了词法和句式等多层次的语言学特征，以实现对原有标注模型的深入改进。通过多组实验论证，结合语言学辅助分析，提出针对性的改进方法从而使模型标注准确率得到了显著提升，证明了结合Max pooling技术的Bi-LSTM语义角色标注模型中融入相关语言学特征能够改进模型标注效果。

Abstract:

With the rapid development of artificial intelligence and Chinese information processing technology, studies relating to natural language processing have reached the level of semantic understanding gradually, while Chinese Semantic Role Labeling is the key technology in the semantic understanding field. Traditional tagging methods depend heavily on the parsing degree of sentence syntax and semantics, so the accuracy of tagging is limited. Aiming at the above problems, this paper improves the basic model of Chinese Semantic Role Labeling based on Bi-LSTM. To solve the above problem, the Max pooling technology is combined in the post-processing stage of the model, and multi-level linguistic features such as lexical item and sentence pattern are integrated into the training to further improve the original annotation model. Through a number of experimental demonstrations, combined with linguistic assistant analysis, targeted improvement methods are proposed to improve the accuracy of model annotation. It is proved that the Bi-LSTM semantic role labeling model combined with Max pooling technology can improve the effect of model annotation by incorporating relevant linguistic features.