Research On Chinese Event Extraction Via Incorporating Attention Mechanism And Long Short-Term Memory Networks

The Internet has become one of the main sources for people to obtain information,and people can easily and quickly get all kinds of information from the Internet.However,with the exponential growth of information on the Internet,the daily information is overwhelming and people are placed in an environment of information explosion.Therefore,how to find the information that people really need from huge amount of information has become an urgent problem to be solved.Consequently information extraction has attracted the attention of many researchers.Event extraction is an extremely important sub-task in the field of information extraction research,as well as one of the puzzles and hotspots in the current information extraction research.The goal of event extraction is to extract the events what people are interested in from the massive amount of unstructured texts that described by natural language(including the type and subtle of the event and the entities,tine and values involved in the event),and then save and present them in a structured way for people to browse.The extracted events can also be used as the input information for machine translation,text retrieval,knowledge maps,recommendation systems,etc.So this research is great practical value and academic significance.This thesis mainly focuses on Chinese event extraction,including Chinese event detection and classification as well as Chinese event argument role extraction.Actually,the event detection and classification task is the process of identifying the event trigger words.This process can be divided into two steps:trigger word recognition and trigger word classification.Combining attention mechanism and long-term and short-term memory neural network,this thesis proposes an ATT-BiLSTM algorithm,which does not rely on part-of-speech tagging and entity recognition.ATT-BiLSTM combines event detection and classification without any manual feature setting.This method can simultaneously capture the local and global information in the corpus.The comparison experiments on the ACE 2005 Chinese event dataset show that,compared with the traditional pattern matching methods,machine learning methods and some existing deep learning methods,ATT-BiLSTM achieves a significant improvement in performance.The event argument role extraction task contains two steps:argument recognition and argument role classification.Combining the information of event triggering words through event detection and classification,this thesis proposes an ATT-DBiLSTM algorithm which combines the attention mechanism and double-layer BiLSTM to complete the extraction of event argument role.By introducing the trigger information,it compensates for the defect that the structural features in the argument extraction are too loose.Compared with other existing methods on the ACE 2005 Chinese event dataset,the proposed ATT-DBiLSTM significantly improves the performance of event argument role extraction.