Research On Weakly Supervised Chinese Relation Extraction

With the continuous development of computer and network technology, a wide range of information appears in front of people。How to extract useful information from vast amounts of information is increasingly becoming an issue of concern. Information Extraction emerges under this background. Entity Relation Extraction is one of the subtasks of information extraction.There are mainly two kinds of approaches for extracting the relations between the named entities. There are Knowledge Engineering Approach and Automatic Training Approach. Knowledge engineering approach has the relative good effect, but it also has obvious shortcomings:(1) The developing of knowledge engineering approach is extremely expensive;(2) It is not flexible. More and more scholars start to devote to automatic training approach research. According to the degree of the manual intervention, automatic training approaches are divided into supervised learning methods, weakly supervised learning methods and unsupervised learning methods.Compared to the precision of supervised learning methods, the precision of weakly supervised learning methods have a big gap. To solve this problem, this article explored a weakly supervised Chinese relation extraction method based on bootstrapping in the current Chinese entity relation extraction research situation and further studied two key components in weakly supervised relation extraction based on bootstrapping. The components are the acquisition of the pattern describing the relationship and the filtration of relational tuples. The main work of this thesis is that the algorithms in the key link of weakly supervised relation extraction based on the bootstrapping were improved and the main issues of the paper as follows:(1) Proposed a the improved method for the acquisition of the pattern describing the relationship-the method of the acquisition of the pattern describing the relationship based on minimum coverage, the patterns got through this method could represent and cover the sentences in the corpus more effectively.(2) Proposed a the improved method for the filtration of relational tuples-the method of the filtration of relational tuples based on mutual assessment, the method could effectively filter relational tuples and improve the accuracy of weakly relation extraction method.Through the above improvements to improve the performance of weakly supervised relation extraction. The proposed method was tested in the open web corpus, and achieved an average accuracy rate of65.6%, which verifies the validity of the methods.