A Research On Entity Relation Extraction Based On Deep Learning

Today's Internet has contained more and more knowledge.By using these knowledge,we can not only construct huge knowledge database,but also improve our intelligent question answering system.Therefore,how to collect and apply these knowledge has become a quite meaningful research field.As we all know,most of the knowledge can be represented by relations between entity pairs.To some extent,mining knowledge from raw texts can sometimes been seen as extracting relations between entity pairs.People has been doing researches on entity relation extraction for many years.However,at the beginning,people tried to solve this problem by rule-based method.This kind of method cost lots of time and human resources.What's worse,the result of this method are not so good.Fortunately,with the development of statistical learning method,people made a breakthrough in this area by using machine learning technology.And with the increasing popularity of deep learning and more and more novel models like RNN and CNN being proposed,the effect of entity relation extraction model has become better and better.This paper talks about how to use state-of-the-art deep learning technology to solve this problem.Firstly we extract four features,word embedding,hypernym embedding,part of speech and relative position.And then we propose two strategies,based on raw text structure and dependency tree structure respectively.The former strategy uses attention mechanism to capture key words and uses CNN model to extract collocation features.While the latter strategy construct CNN model on the dependency path.Compared with the former strategy,this strategy need to do dependency parsing,but in consideration of the small input scale,it has a huge advantage in training speed,and its effect is almost as good as the former strategy.Due to the heterogeneity of these two strategies,we propose two ensemble algorithms to combine them together in order to get better classification results.By using the classical open data set,SemEval-2010 Task8,we find that our best strategy can reach the F score of 85.2%,which shows that combining strategies based on different structures is a quite effective method for entity relation extraction.