Research On Event Detection Based On Recurrent Neural Network

Internet data is growing exponentially.How to extract interesting information from massive data has become a major challenge for human beings.A series of automated processing tasks came into being,including event extraction.The event extraction task is to mine the important information related to the event from the text,and convert the unstructured text into sufficiently simplified structured data.Event extraction is of great significance to some other research fields such as text retrieval,automatic summarization,and text understanding.Event extraction can be divided into two major tasks: event detection and argument role classification.The research work of this thesis is to implement the event detection task based on the recurrent neural network.The event detection task is responsible for finding the event trigger words and corresponding event types in the text,that is,judging whether each word in the sentence is an event trigger word,and if it is a trigger word,it needs to predict the type of event it triggers.The result of the event detection task will directly influence the effect of the entire event extraction task.Therefore,event detection play a very important role in event extraction.The work of this thesis mainly includes:(1)We elaborate on the basic network model referenced in this research.After that,the background knowledge of the two types of recurrent neural network structures and attention mechanism concepts are introduced in detail.Then,We analyze the problems in the traditional model at the word level and sentence level,and propose corresponding improved methods.(2)At the word level,in view of the insufficient attention to the existing event type tags and the label imbalance in multi-classification task,The label information layer se-manticizes the event label information,constructs the semantic connection between the word to be detected and all event labels,and selects the label information with similar semantics to become the auxiliary information of the word,which is of great help for event detection in the case of small samples.The soft gating mechanism scheme is to convert the multi-classification task into a two-stage task which contains one binary classification task followed by fine-grained classification task,so that the categories that the model needs to deal with are relatively balanced.Also the interference of nontrigger words can be eliminated to a certain extent in the fine-grained classification process.The gating mechanism can effectively make a binary classification judgment on whether a word is a trigger word.The improved model is more suitable for short text data sets with fewer data samples.The experimental results show the effectiveness of the model in this thesis.(3)At the sentence level,the "event-event" relationship layer and the "word-entity word" relationship layer are introduced to form sentence-level improved event detection for the shortcomings that the interaction between events within the sentence and the influence of the entity words to the word are not paid enough attention.The "eventevent" relationship layer simulates the possible semantic connection between events in a sentence.It can effectively learn the relationship between events in the same sentence.The "word-entity word" relationship layer uses a customized attention mechanism to simulate the influence of entity words on the detected words to avoid interference from other non-entity words.The improved model is more suitable for long text data sets where multiple events may exist in the sentence.The experimental results prove the effectiveness of the model in this thesis.