Research And Implementation Of Geographical Relationship Extraction System Based On Deep Learning

With the advent of the era of big data,the field of geographic information is undergoing tremendous changes.People's demand for geographic information has gradually changed from a single static geographical location information to a dynamic,diverse geographical information which containing social and human factors.Traditional basic mapping and satellite remote sensing methods are insufficient to meet this change.With the development of machine learning technology,we have a new way to obtain information:through machine learning to analyze network text,identify the semantics of text,and then obtain geographic information in semantics.However,the content of Internet data is heterogeneous and diverse,making the analysis and use of network data very difficult.Knowledge maps have powerful semantic processing capabilities and open organizational capabilities.Therefore,the construction of geographic information knowledge map will promote the development of geographic science and bring more convenience to daily life.At present,the extraction of entity relations is the focus and difficulty of geographic information knowledge mapping research.However,there are currently few studies on entity relationship extraction for geographic information.Due to the lack of data sets in the field of geographic information,research related to machine learning is still stuck in small-scale annotation and simulation data training.The general relationship extraction has made breakthroughs with the development of deep learning,and the algorithms with excellent performance are endless.At present,the extraction of general-purpose domain relationships has made breakthroughs with the development of deep learning,and algorithms with excellent performance emerge one after another.However,due to the lack of data sets in the field of geographic information,there are few studies on the extraction of entity relationships in the field of geographic information.Related research is still at the stage of small-scale labeling and simulation data training.In view of the problems faced by geographic information entity relationship extraction and the current situation of deep learning development,this paper has carried out the following work process.Firstly,this paper calculates the spatial position of the entities in the geographic semantic network and derives the entity relationship triples containing five spatial relationships,and adds them to the original geographic semantic network to enrich the data set content.Based on the remote supervision idea,the geographic information data set is built by aligning the geographic semantic web with the plain text NYT.After that,this paper designs and implements the automatic construction of geographic knowledge mapping system,including network text information capture,named entity recognition,geographic entity relationship extraction and knowledge map data storage.The information is fetched every 30 minutes;the NER of Stanford is used to complete the entity identification and after the entity relationship is extracted,Jena is used to store the finally generated geographic entity relationship triples.For the extraction of geographic entity relationships,this paper first uses word2vec to encode the words of the training set text,and adds the word position information as the model input.Then the residual convolution network is used to extract rich features,and the pooled layer selects the maximum pooling of the segment to further retain the location information.Since the data set assumption based on remote supervision is too positive,a lot of noise will be introduced.Therefore,the whole training process adopts multiple instance learning,and the package is trained as an instance unit.Each instance contained in the package has the same entity pair and relationship label.In the test,the model performed well,and the above-mentioned system for automatically constructing the knowledge map conformed to the design requirements from function to performance.The geographic entity relationship extraction system is tested on the data set constructed above.The experiment shows that when the network depth is 15 layers and the bag size is 200,when the selective attention mechanism is adopted,the model effect is optimal,and the comprehensive accuracy rate is 74%,higher than the PCNNs model to test the effect on this data set.The above-mentioned system for automatically building knowledge maps meets the design requirements from function to performance.