Causality Extraction Of News Events

Natural language is full of cause-and-effect expressions.Every phenomenon or thing has its cause,and people use cause-and-effect relationships to describe the laws of movement and development of things.At present,a mass of news data exists on the Internet,which involves massive knowledge related to events.Identifying the causal relationship between events can help to excavate the logical knowledge contained in news and depict the development logic between events.Based on the above statements,this paper focuses on the causal extraction of events in the news and constructing the event logic graph.For the event detection task,previous research works focused on improving the model's ability to model text without effectively using the event trigger words which can indicate relevant events.Therefore,the first work in this paper proposes a trigger word extension mechanism,which extends and constrains the trigger words in terms of syntactic coherence and semantic similarity,respectively.The experimental results show the method in this chapter can obtain more event-related trigger words with less training data.The performance of event detection can be improved by adding extended trigger word information.On this basis,in order to identify causal relationships between events and obtain more generalized event representations,the second work in this paper focuses on event-based causal extraction.In this paper,cause and effect are defined as consisting of a noun and its predicate or state.The causal event arguments as well as the causal trigger word need to be extracted.Based on the extension of causal trigger words,we use a double pointer labeling network and incorporate a negative sampling mechanism for the extraction of causal arguments.Besides,we propose an argument co-occurrence module to assemble different arguments to form structured events.The effectiveness of the model is proved by the results of ablation experiments on Tong Hua Shun causality dataset.Based on the causal extraction model,the third work performs causal extraction on large-scale news data.Here we extract and obtain about 6 million pairs of causal relationships,and add processes such as event generalization and event fusion to construct the event logic graph mainly in the financial domain.Focusing on event-based causal extraction,this paper sequentially investigates events detection,the extraction of causal event relations,and the prediction of causal events on large-scale news to finally construct the event logic graph in financial domain.