Research On Event Detection Based On Knowledge Expansion And Information Fusion

Event extraction aims to identify specific event information from text.It's a critical task in the Information Extraction field.Automatic Content Extraction(ACE)defines event extraction as:trigger identification,event type classification,argument identification,and event role classification.This paper focuses on the first two subtasks,which is referred to as "Event Detection".ACE05 corpus labels the triggers and corresponding event types for sentences.Currently,the performance of trigger identification and event type classification has reached 70%or more.However,the event detection is difficult due to the sparse data and the unbalanced distribution of data in ACE05 corpus.Additionally,the ambiguity caused by the polysemy of triggers also limits the accuracy of event detection.In order to solve the problems of sparse data,unbalanced samples distribution and triggers ambiguity in the event detection task,this paper proposes novel event detection method based on knowledge expansion and information fusion utilizing the neural network.The research content specifically includes the following three aspects:(1)Training Set Extension Based on Frame Semantics for Supervised Event De-tectionThe sparse data and unbalanced samples distribution of ACE05 corpus cause insuf-ficiently training of event detection model.Thus,this paper proposes to expand ACE05 corpus.The define of frame is similar to event,and the FrameNet contains rich labelled samples,which provides the possibility for corpus expansion.We utilize the relevance of the lexical units with triggers to obtain the expanded data.Then,we distill the data via the similarity between the event type and the frame type.Finally,we get suitable expansion samples.We train different event detection models with expanded corpus.The experimental results show that our approach can improve performance of the event detection.(2)Improving Event Detection with Entity Relationship via Interactive LearningTriggers can express different semantics with different contexts and trigger different event types.In the event type classification subtask,ambiguous triggers will mislead the model and cause lower performance.We find that the entity relationships in a sentence can constraint the semantic of triggers,and alleviate ambiguity problem.To this end,this paper takes the entity relationships as feature,and designs the event detection model based on the attention mechanism,to fuse relationship information and capture key contextual information.In addition,this paper utilizes the interactive learning to joint extract event and relation.Both modules use outputs of each other as their own features.Experiments verify that our method makes the F1 score of trigger detection and event type identification reach 75.3%and 72.7%.The performance of relation extraction also gets better.(3)Event Detection Based on Attention-guided Heterogeneous Graph Convolu-tional NetworkTo integrate the syntactic and semantic information of the sentence,this paper propos-es the event detection method based on attention-guided heterogeneous graph convolutional network.This paper utilizes the entity relationships and dependency parser of the sentence to construct a heterogeneous graph,and assigns different weights to each paths in the graph via node-based attention mechanism.The multi-layer graph convolution operation is used to capture syntactic feature and semantic feature.This method fuses the relationship infor-mation and dependency information in the sentence,and strengthens the information inter-action between the context and triggers.Experiments show that this model can improve the performance of the event detection effectively.Based on knowledge expansion and information fusion,this paper alleviates data s-parsity,unbalanced data distribution,and triggers ambiguity problems based on different external information and suitable neural network model.The performance of event detec-tion also reaches the optimal.The F1 score of the trigger recognition reaches 76.8%,and the F1 score of the event type classification reaches 75.2%.