Research On Entity Relation Extraction Algorithm Based On Semi-supervised Machine Learning

With the rapid development of the Internet technology,we have been everywhere filled with information,people urgently need to get valuable information from the massive data,information extraction technology came into being.Entity relation extraction is one of the important branches of information extraction,it can extract the entity,the attribute value and their relation from the text,so as to transform the unstructured data into the structured data,thus the entity relation extraction technology has received great attention.At present,the entity relation extraction technology is mainly combined with the machine learning technology to complete the task of relation extraction,this thesis will focus on the study of entity relation extraction by the semi-supervised machine learning method.In the semi-supervised machine learning algorithm Bootstrapping algorithm,how to suppress semantic drift is one of the main challenges.In this thesis,we propose a novel method to select the trigger words,and define a flexible measure of the trigger force.The experimental results show that the proposed method makes the semantic constraint of the pattern stronger,so that the recall rate is greatly improved.In addition,it is necessary to add the new relation pattern to the initial seed set in the semi-supervised algorithm,which requires us to define a reliable pattern similarity measure method to select the reliable relation patterns to constantly extend the seed set.In this thesis,we propose a simple and effective kernel function to measure the similarity between a new pattern and the patterns in old seed set.The experimental results in the thesis show that the overall accuracy of the method is increasing with the expansion of the seed set.Finally,in order to make the newly added relation pattern more reliable,this paper proposes a model with two kinds of machine learning algorithms,so that the confidence of newly added relation patterns can be improved.The experimental results show that the model can obtain a more stable and effective prediction effect.