Study On Event Extraction Model Construction For Emergency Case Base

Recent years, the worldwide frequently-occurred emergencies have been a great threat for the security and development of human society, which has caused great concerns of scholars and governments at home and abroad. Due to its suddenness, emergency is lack of information, which results in insufficient decision criteria, increasing risk for decision making and more difficulty for event handling compared with common events. Extracting structured information from emergencies-related information and constructing emergency typical case base are effective for improving the ability to cope with emergencies. Thus, constructing event extraction model for emergency case-base to realize automated case information extraction is of great significance for emergency decision making.In order to make our model better serve the emergency case base, firstly the paper defines concepts involved in the extraction process such as emergency case information, emergency and emergency extraction framework, then a quaternary feature vector is presented describing emergency cases from four aspects:case type, case description, solution, decision-making environment and defined different event elements under each feature dimension according to different types of emergencies. Meanwhile, emergency case information extraction framework with inheritance mechanism is constructed based on this feature vector model. Then, for different aspects features, aspects of the framework are divided into three categories:event time, event location, and other aspects. Extraction methods and patterns of each category are discussed in detail as follows:1) Time category. The concept of basic time element is put forward to refer to the forms and features of time expressions in emergency related documents, in representative of basic components in time expressions. Matching patterns and merging rules of basic time element are also summarized and explained. Putting forward non-numeric time numerical value strategies, time chunk and fuzzy time numerical value converting rules for the standardization of time expressions are discussed. Event time is inferred with document location and importance of time expressions.2) Location category. Additional domain dictionary and location dictionary are constructed to solve the misrecognition of domain phrases and low recognition rate of foreign places in segmented documents; the concept of geographic elements is put forward for the low recognition rate of place chunks, extraction pattern and merging rules of geographic elements are also summarized; place database, place name-abbreviation chart and location co-reference pro-form resolution strategy are constructed for the standardization of place expressions; event location is inferred with place document location, importance, place relations of place expressions.3) Other aspects. Other aspects of an event are converted into meta-events according to relations between event other aspects and meta-events, types of meta-events to be extracted are also defined. Candidate events extraction methods are computed using expanded trigger set and HowNet-based similarity. Different aspects are extracted with pattern based and semantic role labeling based methods according to feature of different aspects.4) Events merging. Identifying and merging rules using event location and event time as feature are put forward to deal with multiple document environment.Experiments on public sanitation emergency audio material shows that our extracting model is feasible, which achieves a high precision and recall rate.