Research On Slot Filling Method Based Upon Entity Relevance And Semantic Information

Given an entity query,slot filling aims at extracting its attribute values from a largescale corpus.This research includes two key tasks: 1)Source information(relevant documents)acquisition,the task is to retrieve relevant documents with the query entity and the reference document from a large-scale corpus,the relevant documents are regard as source information which not only contain the query entity but also provide significant contextual information.2)Attribute extraction and inspection,the task is to extract the candidate attribute values of query entity from the relevant documents and inspect the correctness of the candidate attribute values to form a reliable result.The researches of this paper are based on the two tasks,we apply both the semantic information and the entity relevance information to the source information acquisition,and attempt to apply them to the attribute extraction process independently.All of this form a slot filling model which congregate retrieval and extraction technology.In general,the content of this paper are as follows:1)Document retrieval method based on entity supervision and topic model: This paper builds the source information acquisition model based on the retrieval framework,and focuses on solving the problem of the error of the related document acquisition caused by the ambiguity of the named entity.This paper proposes a retrieval method that combines topic semantics and entity communities,and uses pseudo-related feedback to enrich the background information of the entity query.In the retrieval process,this paper uses the neural network model to combine topic semantics and entity community features in the calculation of the relevance degree.Experimental results show that this method effectively improves the performance of document retrieval(F1 measure increased by 6.4%).2)Attention-based CNN for entity relation classification: This research deals with sentences one by one in related documents,aiming to use the classification model to identify candidate attribute values for specific attribute.However,there are often many kinds of semantic relations in the sentences to be classified,which misleads the classification of the semantic relation between the query entity and the candidate attribute values.In order to solve this problem,we propose an attention-based convolutional neural network entity relation classification model,which uses attention mechanism to give higher attention to the semantic information closely related to the query entity,so as to optimize the relation classification model.Experimental results show that this method obtains better performance than the current entity relation classification model(F1 measure increased by 2.94%).3)Unsupervised slot filler refinement via entity community construction: Identifying the attribute values of query entity with the relation classification strategy helps to discover multiple textual representations of the attribute value,but it often contains redundant and incorrect attributes.Existing studies have attempted to filter redundancy and error attributes with confidence,but there is still a large gap in performance.To solve this problem,we propose unsupervised slot filler refinement via entity community construction.Use hierarchical clustering to generate multiple communities of entity attribute values,then use the entity graph model to mine the community related to the query entity,and finally filter the wrong entity attribute values with the related community of the query entity.Experimental results show that this method can effectively improve the performance of the slot filling system(F1 measure increased by 6.8%).