Research On Chinese Event Extraction

In recent years,with the rapid development of information technology,massive information is presented in the form of text.In order to deal with the challenges brought by the explosion of information,people urgently need a technology to extract structured information from unstructured texts,information extraction technology arises at the historic moment.Event extraction is one of the most important sub tasks in information extraction,it is also the focus of current research.Event extraction finds events and arguments from the natural language text and saves the event into the knowledge base for the use of the information query.Event extraction has important significance for machine translation,text retrieval,automatic text summarization.The current event extraction technology is mainly based on the supervised method.In this paper,a series of Chinese event extraction techniques are studied on existing supervised event extraction methods.At first,we proposed a Chinese trigger extraction method based on event semantics.The concept of event semantics is proposed,and event semantics is explained in three aspects:trigger word,event arguments,trigger-argument relation.In order to solve the problem of event related argument extraction,an algorithm based on the combination of dependency parsing and heuristic rules is proposed.With the addition of event semantics on the baseline system,the performance of the Chinese event extraction is effectively improved.Then,we propose a Chinese trigger extraction method based on the discourse consistency in triggers and arguments.According to the frequent missing of arguments in Chinese corpus and the strong correlation between triggers and arguments,we divide discourse consistency into two parts,trigger consistency and argument consistency.Trigger consistency can also be divided into three parts,same events,similar events and co-occurrence events.We train a discourse-level classifier to assist the baseline,which is a sentence-level system.The result on ACE corpus shows that the performance is improved.Furthermore,we propose a Chinese event extraction method based on the joint model.Traditional methods rely on pipeline with multiple stages,which suffers from the error propagation and mission splitting,in this paper,a structured perceptron model is introduced into Chinese event extraction.In the decoding stage,beam-search algorithm is used to reduce the space search cost.The average perceptron is used in the model to prevent the over fitting.The sentence-level and discourse-level features on the joint model are also studied.The experimental results show that the effect of the Chinese event extraction system based on the joint model is better than the baseline system.At last,we propose a Chinese trigger extraction method based on neural network.Aiming at solving the problem of error propagation and Chinese language specific issues in traditional methods,a neural network model is proposed to Chinese trigger extraction.We use Bi-LSTM to get sentence-level features,CNN to get lexical-level features,and use CRF tagging model to annotate characters.Experimental results on ACE2005 Chinese corpus show that the proposed neural network method is superior to the traditional methods based on NLP tools and feature engineering.