| A fundamental goal of scientific research is to understand the causal relationships between things,to explore the nature behind phenomena and the underlying laws of how things develop.However,existing approaches to AI research are more often based on a purely data correlation perspective rather than causal reasoning.In order to better understand causal relationships and advance causal reasoning,the most important step is to extract causal information from real-world data of all kinds.Therefore,this paper focuses on polymorphic causality extraction from natural language text,chapter-level event causality identification and inference,and ultimately the construction of a causal knowledge graph covering multiple domains.First,this paper focuses on polymorphic causality extraction,which consists of two tasks: intra-sentence causality extraction and inter-sentence causality identification.For the intra-sentence causality extraction task,this paper adopts a pipelined approach:all noun words or phrases(Nominals)in a sentence are extracted first,and then the semantic information of the causal signal words in the sentence is used to identify the causally related Nominals pairs.This approach avoids the problem of causal nesting that exists in existing work that directly employs a sequence annotation approach.For the inter-sentence causality recognition task,the performance of implicit causality recognition is poor due to the lack of causal signal words.In this paper,we adopt the idea of cue learning to transform the inter-sentence causality recognition task into a form closer to the pre-training task of large-scale language models,so as to fully exploit the common sense knowledge embedded in large-scale pre-training language models and effectively improve the performance of the model on the implicit causality recognition task.Secondly,this paper investigates chapter-level event causality recognition,mainly including intra-sentence and cross-sentence event causality recognition.To address the problem of insufficient annotation data in existing research work,this paper proposes a method of incorporating common sense knowledge from external knowledge bases into the coded representation of events to help the model identify causal relationships between events.At the same time,in the cross-sentence event causality recognition task,the large span of sentences and the complexity and variety of contextual information lead to the problem that the model cannot effectively perform causality recognition.In this paper,the inter-sentence causality recognition study in the previous chapter filtered out a large number of negative sample pairs,thus improving the accuracy of the model recognition.Based on the chapter-level event causality recognition,this paper further explores event causality inference,trying to reason out what the corresponding result(or cause)event is given the context and the cause(or result)event.Finally,this paper uses the above research methods to perform causal relationship extraction on textual data such as news from the football domain,news from the financial domain and diagnostic reports from the medical domain.The graph database Neo4 j is then used to construct a multi-domain causal knowledge graph and to explore the application of causal knowledge graphs on the corresponding domains. |
