Temporal Relation Extraction Of Event Centric Knowledge Graph In Tourism

In the period of " 14th Five-Year Plan",Our country will enter the era of mass tourism.With the increasing tourism users,providing intelligent and personalized information services for users has become an urgent problem for tourism development.At present,the Internet is filled with large amounts of tourist spatial-temporal trajectory information,but lacks effective mining and utilization.Therefore,efficient information management and use become the key to implement smart tourism.In this thesis,we construct tourism event-centric knowledge graph to organize and manage spatial-temporal information.In addition,tourism event-centric knowledge graph is beneficial for the development of downstream applications,such as:question answering,POI(Point of Interest)recommendation.However,event temporal relation extraction,which is an important component of event knowledge graph,are rarely studied.In this thesis,we firstly study event temporal relation extraction in English,and then try it in Chinese tourism data to construct tourism event knowledge graph.Compared with the rule-based methods,tourism event-centric knowledge graph has more accurate temporal relations and improves the reliability of the knowledge.In order to verify the effectiveness of tourism event-centric knowledge graph,we study the downstream application of knowledge graph,POI recommendation.We propose a personalized recommendation algorithm,Event Knowledge Graph Attention Network for Augmenting Sequential Recommendation.The main work and innovations of this thesis are as follows:Firstly,we propose a cross-sentence temporal relation extraction model.At present,the method of event temporal relation extraction based on pre-trained language models have insufficient performance in cross-sentence temporal relation extraction.To solve these problems,we propose a sentence pair temporal ordering task to model the temporal order of two consecutive sentences and a multi-task learning framework is employed to integrate the sentence pair temporal ordering task with a temporal relation classifier to improve the performance of cross-sentence temporal relation extraction.In addition,we propose a Vague relation subsampling method to deal with the class-imbalanced data on the small datasets.Experiment results show our model not only outperforms SOTA methods on benchmarks,but also improves cross-sentence temporal relation extraction performance.Then,we propose personalized recommendation algorithm,Event Knowledge Graph Attention Network for Augmenting Sequential Recommendation.The data in Tourism POI recommendation is sparse and the performance of this method is poor.To solve these problems,we improve the performance based on deep structured semantic model framework.In the item representation,we aggregate the k-hop neighbors of entities in the tourism event knowledge graph through the graph attention network.In the user representation,we use a short sequence augment algorithm based on the pre-trained model for sparse data.Then,based on the prediction results of the recommendation model,we propose a reranking algorithm to optimize results.Finally,we construct a smart tourism platform for the research and development of related application to lay the foundations for future development.