Research On Event Detection Without Trigger Word Combined With Tag Information

With the development of the Internet and the exponential growth of the amount of information,how to obtain the required information from a large amount of information is an important problem.Therefore,the research on the related technology of information extraction is becoming more and more important.Event detection is a sub task of event extraction task in natural language processing,and it is an important direction of information extraction.Given a piece of text,event detection aims to detect potential event types.At present,the mainstream of event detection is to identify the trigger word in the text and judge which event type the trigger word belongs to.This event detection method based on trigger words will increase the cost of data annotation.In addition,trigger words are not essential in the event detection task.Therefore,taking the event detection method without trigger words as the basic idea,this thesis studies how to realize the event detection of text without trigger words.Specifically,this thesis analyzes the problems of sentence level and document level event detection tasks,and puts forward the corresponding algorithms respectively.1)Sentence level event detection without trigger words.Different from other fields,tags in natural language field often have semantics,but at present,most studies,including event detection,often use tags as a unique hot vector.Therefore,this thesis proposes an event detection algorithm combining label perception and multi task learning.On the one hand,the text to be detected and the label text are represented as vectors,and the two vectors are used to calculate the probability that the text to be detected contains the event type corresponding to the label;On the other hand,an auxiliary task is constructed for the event detection task to enhance the generalization ability of the model through multi task learning.Experiments on ace2005 dataset show that the proposed algorithm F1 reaches 73.4%,which is significantly better than other non trigger word event detection methods.2)Document level event detection without trigger words.The previous event detection work was mainly carried out on a single sentence,and there was little event detection work for documents.Different from ordinary sentence oriented event detection tasks,document level event detection tasks need to focus on the situation where there are multiple events,because a sentence often contains only one event,but a document is likely to contain multiple events.Aiming at the problem of multi event correlation in document level event detection task,this thesis proposes a neural network model based on Bi LSTM and multi type attention mechanism.The model extracts the local and global information of documents through Bi LSTM and self attention mechanism,compresses the word vector into document vector with the help of word level and sentence level multi head aggregation attention mechanism,and completes document level event detection by multi label classification.The multi head aggregation attention mechanism used in this model can be effective in multi event scenarios.Due to the lack of document level event detection data set,this thesis constructs a document level event detection data set based on marine news,and makes an experimental analysis on the proposed document level event detection model and other baseline models on the data set.The experimental results show the effectiveness of the proposed model.