Research On Key Technologies Of Event Causality Graph Construction For Text Data

The event logic graph describes the patterns and regularities between events.It has significant application value in artificial intelligence tasks such as question and answer systems,intelligence analysis,event prediction,and decision support.Currently,due to the concern for abstract and generalized events in the event logic graph,which encompasses various logical evolution relationships between events,it is unable to meet the demands of some real-world scenarios that emphasize the concretization of events,the richness of examples,and the focus on causalities between events.Existing event extraction and relation extraction models cannot satisfy the needs of event causality graph construction which emphasizes event concretization and the causality between events.Therefore,this dissertation focuses on three key technologies in event causality graph constructions for text data: document-level event argument extraction,event causality identification,and event graph representation.It conducts in-depth research from four aspects: role knowledge prompting,knowledge interaction graph guidance,event relationship data augmentation,and the integration of knowledge graphs.The main studies and contributions of this dissertation are:1.Document-level event argument extraction aims to extract event arguments from the entire document.It is the foundation for event causality graph construction,providing abundant and complete elements of event instances.To address the issues of scattered event arguments,overlapping arguments and roles in document-level event argument extraction,a model called RKDE(Role Knowledge Prompting for Document-Level Event Argument Extraction)is introduced.This model frames the task of document-level event argument extraction as a slot-filling task that simultaneously extracts all arguments.Two prompting learning paradigms are introduced to enhance the model’s semantic understanding of the text.One is role definition template which prompts the model to focus on important information in the text,and the other is event description template which narratively describes the argument roles.To improve the interaction of templates and role information,a guiding mechanism is employed,providing precise prompts to the pre-trained language model,and then the generative pre-trained model is used to simultaneously generate all arguments through a textspan query.Experimental results show that,compared to existing document-level argument extraction models,the proposed model is not only able to extract scattered event arguments simultaneously but also demonstrates superiority in the extraction of overlapping arguments and roles.On two widely used document-level datasets,RAMS and WIKIEVENTS,the F1 scores for event argument classification have improved by 3.2% and 1.4% respectively.2.Event causality identification aims to recognize whether there is a causality between different events in the text,which is crucial for event causality graph construction and provides accurate knowledge of causal logic.To address the challenges of identifying implicit causalities and insufficient interaction between events and knowledge in event causality identification,a knowledge interaction graph guided event causality identification model called KIGP(Knowledge Interaction Graph guided Prompt Tuning for Event Causality Identification)is proposed.This model incorporates external event knowledge and designs event knowledge interaction graphs based on the perspective of text,event,and knowledge interaction.By constructing an event knowledge interaction graph,it uses a graph convolutional neural network to aggregate event vectors with knowledge.In order to activate the causality recognition ability of the pre-trained model,a prompt template with causal event extraction task is designed,and the interaction guidance mechanism is utilized to enhance the representation of events in the prompt template and capture the deep semantic relationship of the document.Further the identification of complex implicit causalities between events is improved.Experimental results show that on two public event causality identification datasets,Event Story Line and Causal-Time Bank,performance metrics have significantly improved.Compared to the recent graph structure ERGO method,the model’s F1 scores have increased by 6.3% and 2.9% respectively.3.Most existing methods for event causality identification and extraction rely on supervised learning models that require a substantial amount of training data.However,due to the lack of a unified framework for defining event relationships,the scale of event relationship datasets is generally small.To address this issue,a novel method of event relation data augmentation based on relation prediction called ERDAP is proposed.This method views event relation data augmentation as a relation prediction task and uses an event relation graph convolutional neural network to predict event relations.The high-quality event relation triples generated are used as new training data to expand the event relation texts.It adopts an end-to-end approach to encode events at the encoder end,generate potential hidden feature vector representations of target events through the event information in the constructed event relation graph,and predict event relations at the decoder end.Based on the prediction outcomes,new training data is produced,serving as a form of data augmentation.Experimentation reveals that the enriched relation data boosts the quantity and variety of the dataset,providing a richer corpus for model training.Using Att-BiGRU and KIGP models to extract event causalities from the CEC dataset before and after data augmentation,their performances have improved by 1.4% and 1.8% respectively.4.Event graph representation is the process of transforming extracted event information and identified causalities into a graphical structure,providing a powerful tool for understanding and analyzing complex sequences of events.In response to the lack of interaction between events and entities in existing event graph representation models,and the inability to adaptively represent dynamic knowledge,a multi-level event representation model that integrates knowledge graphs,termed MEMK(Multi-level Event Representation Model Integrating Knowledge Graph)is proposed.This model designs and constructs an event semantic representation model by combining dynamic event knowledge and static data knowledge from the perspective of fusing knowledge graphs.Event type expansion is realized based on large language model retrieval-augmented generation.Additionally,it adopts a combination of the knowledge graph resource description framework and the property graph model,leveraging property-associated knowledge graph entities to match corresponding contextual knowledge,forming a pathway from event knowledge to entity knowledge,and thus achieving dynamic representation across different search spaces.Finally,based on the proposed model,Chinese event causality graphs for emergency incidents and military exercises are designed and constructed.Experimental results show that compared to relational storage,property graph storage improves retrieval efficiency by 3-4 times when the query depth is 4-5.Additionally,the event causality graph can provide effective support for revealing the evolution patterns of events and mining the motives and transmission paths of events.