A Study On Fine-Grained Spatial Relation Extraction Technology

Natural language is an important way for humans to describe spatial information.The text data contains rich spatial information and extracting useful spatial information from text efficiently and accurately is beneficial for many related applications such as Question Answering System,Knowledge Graph Construction,Geographic Information System,etc.ISO-Space is currently a relatively mature text-oriented spatial information annotation specification,which defines the concepts of spatial elements and spatial relations.Spatial elements are the basic semantic units,and spatial relations describe the semantic associations among spatial elements.Human languages exhibit a variety of strategies for communicating spatial information,which brings a great challenge to spatial information extraction tasks.Based on ISO-Space,this thesis studies the text-oriented finegrained spatial relation extraction technology,and the main research work includes the following aspects:This thesis studies the fine-grained spatial relation extraction task,which aims to extract spatial relations and fine-grained type attributes from text data with annotated spatial elements.For this task,this thesis proposes a spatial relation extraction approach called DI4 SR,which consists of two stages: firstly,semantic dependencies between pairs of spatial elements are identified by a dependency identification model based on a deep self-attention neural network,and then semantic dependencies are combined into complete spatial relations according to predefined rules.DI4 SR performs satisfactorily on the Space Eval benchmark.DI4 SR approach fails to make full use of all role information in the spatial relation when identifying fine-grained type attributes.Therefore,this thesis proposes a fine-grained spatial relation classification approach called FGSRC,which identifies fine-grained type attributes for complete spatial relations.Experimental results show that the DI4SR+FGSRC approach outperforms the single DI4 SR approach.This thesis also studies the end-to-end fine-grained spatial relation extraction task,which directly extracts fine-grained spatial relations from the unannotated text.For this task,this thesis proposes a joint extraction framework based on multi-task learning to effectively utilize the potential relevant information among related tasks.Experimental results show that the joint extraction framework based on multi-task outperforms the traditional pipeline extraction framework.