The Research On Event Relation Strength Calculation Method Merged Sentential Semantic Feature

With the popularization of the Internet around the world, the speed development ofnetwork transmission technology, the rapid growth of web data and the swelling networkdata, we are led into the era of "information explosion". Among the information that peopleassociate with in their daily life, the vast majority of the information is in unstructured textform. Therefore, how to make an accurate structural analysis of text and transform it into awell-understood form is one of the issues that the current text analysis are facing. Events asthe basic unit of human cognition, understanding and memory, plays an important role inunderstanding and memorizing text. In higher granularity, text is a description of event sets.Generally, the information associated with an event topic is scattered and isolated indifferent times and different places. Thus this paper focuses on the events and coverts textinto events and the relation strength among them. The proposed method keeps semanticinformation in text fairly well and gives a form of the semantic information for people tounderstand and memorize.Isolated trigger and event argument can only provide partial information of the eventfor person, which can not satisfy the form requirements for understanding and memorizing.Event extraction researches focus on recognition of event trigger or event argument. Inorder to extract the complete event, a new method merging sentential semantic features isproposed for event extraction in this paper. The method takes words as the sequencelabeling units, and trains the conditional random field model to recognize the trigger word,and applies a rule-based post-processing to correct the triggers. Then the dual-layerconditional random field model and sentential semantic features are used to identify theevent argument. Finally the shallow layer location position information is used fordistinguishing the first two parts recognition results which represent the same event, and inthe guidance of event mapping mode, sentential semantic components are extracted usingsentential semantic information. The experimental results show that, the average F-score ofoverall results reaches59.9%, the introduction of sentential semantic feature can extractevent within sentence effectively, which provides a new thought for event extraction.For that researches on events relation recognition only focus on identification ofcertain event type, this paper starts from event relation quantification, and presents amethod for events relation strength computation in order to cover a wider range of events relation. First of all, typical words are selected from event to represent event vector, and thevector is optimized with context information and core event information. Then the text isconverted into a word-event matrix namely the representing matrix and tf-idf weighting isapplied to evaluate how important a word is to a document. LSA is used for analyzing therepresenting matrix and cosine similarity is applied to compute strength between a pair ofevent-event. Finally, the pair event whose strength is greater than the threshold is selectedto construct text representation. The method is evaluated by the automatic textsummarization experiment. The results show that when the compression ratio is30%, theaverage accuracy rate reaches64%, indicating the method can effectively calculate eventsrelation strength, and provides the reference value for the application based on eventsrelation strength.