Research On Event Detection Based On Self-attention Neural Network

In recent years,with the rapid development of Internet and the explosive growth of text information,the problem of “information overload”has become more and more serious.How to quickly and accurately extract interesting information for users from a large amount of unstructured information has become a hot issue in the field of natural language processing.Therefore,information extraction technology has received more and more attention.As an important subtask in information extraction,event extraction is designed to extract specified event information from natural language text and store it in a structured form.The event extraction is divided into two subtasks: event detection and event argument extraction.This paper focuses on the first subtask.Considering that event detection still faces the following challenges: a)the same type of event may be triggered by different trigger words,which corresponds to the rare trigger word problem under the same event type;b)the same trigger word may trigger different event type,which corresponds to the ambiguity of the event type of the same trigger word.Thus,this paper proposes the following two research methods based on the ACE 2005 corpus:(1)Event detection based on the label-aware dual self-attention network.From the representation perspective,the distribution of the trigger word and the event type label is consistent in the high-dimensional semantic space,indicating that the label information can capture the type-specific context,which is useful for the detection of rare trigger words.In addition,direct associations between any words may provide useful clues.However,previous work either introduces some external event instances with the help of the knowledge base or designs complicated reasoning rules to mine crossentity\event\sentence information to alleviate data sparsity,which completely ignore the semantics of labels information and are difficult to cover all the semantic rules.Therefore,this paper proposes to use the semantic information of the label space to explicitly enhance the information representation in the word space,which simply and efficiently solves the rare trigger word problem to some extent.(2)Event detection based on the hierarchical topic-driven self-attention mechanism.Some work tends to use a broader context to provide useful clues for disambiguation.They can be divided into two categories: 1)feature-based methods;2)representationbased methods.The former capture cross-sentence clues by devising complex inference rules,which not only need to elaborately design rich features,but also are difficult to cover all semantic rules.While the latter obtain a global document representation through an unsupervised or supervised method.Although this can indeed bring valuable disambiguation evidence for ambiguous words,it may also introduce some noise information,which would exacerbate word confusion and even interfere with the detection.Therefore,this paper proposes to use the neural variational inference topic model to obtain common global document information and specific trigger meaning clues to reduce the ambiguity of the trigger words.In the ACE 2005 English dataset,the effectiveness of the proposed method is demonstrated by comparison with the experimental results of the previous methods.