Research On Domain-Specific Event Extraction

Current studies on event extraction mainly focused on general domain,while domainspecific event extraction is lack of dedicated corpus and targeted extraction methods.It's necessary to define a set of the appreciate event types and to provide effective extraction methods for domain-specific event extraction.This paper mainly focuses on the in-depth study of event extraction techniques in the field of military intelligence,which includes the following three aspects:(1)This dissertation proposes an event representation system for the domain of military intelligence and then builds an event corpus.First,it defines a domain-specific event representation system,which includes 15 types of subjective events and 21 types of objective events.Besides,we propose the concept of "non-entity argument" and defines the argument extraction task as two parts: entity-based argument extraction task and non-entity element extraction task.Finally,we annotate a domain-specific event corpus,which contains 531 articles and 5624 events,which achieves a good consistency in annotation check.(2)This dissertation proposes a method of domain-specific event extraction on the characteristics of local entities.First,it implements a traditional feature-based trigger extraction system as the baseline.Then,it proposes a trigger extraction method on the characteristics of local entities.This method uses the pre-screening entities and CNN model to locate the entities that are most important for trigger recognition,and then uses the combination information of the entity and the trigger to identify the trigger.The experimental results show that our method can reduce the interference of unrelated entities in trigger extraction and outperforms the baseline by 1.7 in F1-score.(3)This dissertation proposes a non-entity argument extraction method on structural representation.Inspired by the methods of negation and uncertainty scope detection,it uses syntactic subtrees as classification units,constructs a variety of plant features and structure features of candidate subtrees to explore its syntactic relations with trigger.And then it uses a tree kernel-based classifier and a neutral network classifier to identify the non-entity arguments in candidate subtrees.The experimental results show that our method outperforms the baseline in PCS by 13.5%.This result ensures that the identified non-entity arguments are complete syntactic structures,eliminate the influence of the complex internal structure of non-entity arguments on recognition.