Research On Spatial Information Extraction From Text

There is a lot of spatial information in texts,extracting the spatial information from texts is beneficial to many natural language processing tasks,such as named entity recognition,question answering,spatial reasoning,natural language understanding and so on.ISO-Space is a specification about the spatial information,including the spatial elements and spatial relations.Spatial elements are the basic unit of describing spatial information,while the spatial relations catch the relevance between spatial elements.This thesis applies deep learning technology to spatial information extraction,s-tudying the spatial elements recognition and spatial relation extraction according to the ISO-Space specification,the main work and contributions of this thesis are as follows:1.For the spatial elements recognition,we consider it as the sequence tagging prob-lem and propose a BiLSTM-CNN based hybrid method.This method introduces the domain specific word embedding and char representation to the embedding layer,which emphasizes the characteristic of spatial lexical;In the step of feature repre-sentation,this method uses the BiLSTM and CNN for sentence representation.At last,we use the conditional random field for tag reasoning.This method gets great performance in the SpaveEval spatial elements recognition task.2.As the lack of labeled data on spatial information,we propose a BERT based method for spatial elements recognition.BERT facilitates pre-training deep bidirectional representations on large-scale unlabeled corpora,and contains rich semantic infor-mation,also we utilize conditional random field for tag reasoning after BERT pre-training model.This method achieves better result than the BiLSTM-CNN based hybrid method.3.For the spatial relations extraction,we split the problem into two stages:role po-sitioning and relation recognition.In role positioning stage,we generate candidate relations according to ISO-Space specification and some other heuristic strategies.In relation recognition stage,we consider it as a classification problem,and pro-pose a CNN based method.This method utilizes the CNN for feature extraction and combines the position features to judge whether the candidate relation is valid.This method obtains the best performance in SpaceEval spatial relation extraction task.