| With the rapid development of big data,the Internet is filled with a huge amount of text information.As an important step of event extraction,event detection is to obtain the type of events contained in the text,which has certain guiding significance for the task of subsequent event extraction.However,because Chinese text does not have a space between similar English words as a natural separator,it is often necessary to additionally break the Chinese text before subsequent natural language processing tasks can be performed,and word segmentation errors are often a major obstacle affecting event detection performance.In this paper,a Chinese event detection model based on mixed representation is proposed,which uses the skip-gram model in Word2 Vec to obtain the character embedding of text;An external lexicon is introduced through Lattice-LSTM,blending character-level and word-level features;the attention mechanism is used to give higher weight to the trigger words in the sentence;and finally the word information and sentence information are jointly encoded using Lattice-LSTM,and the event types contained in the sentence are classified by softmax.Experimental results on the Du EE1.0 dataset show that the Chinese event detection model based on the mixed representation proposed in this paper has significantly improved the fusion effect of information compared with other methods,and the F1 value reaches 84.28%.This paper also proposes an event detection model that leverages global information enhancement to solve the problem of multiple meanings of trigger words.This paper uses Lattice-LSTM to fuse external lexicon to enhance the representation of word and character information in the text;sentence-level encoding through Bi-GRU while capturing the semantic information of the context;using the attention mechanism to assign higher weight to important sentences in the document as a global information guidance model;then the words,words,sentences,and document information are stitched together as input to the encoding layer,and then the text features after the fusion are encoded by Lattice-LSTM.Finally,the event type is obtained by the softmax layer.Experimental results show that the event detection model proposed in this paper using global information enhancement has improved the precision of the Du EE-fin dataset by 1.66% compared with the TRNN model,and achieved a relatively satisfactory effect. |
