Research On The Construction Of Geographic Entity Relationship Based On Neural Network

Geographic text data is readily available,large-scale,and contains a large amount of geospatial knowledge.However,the traditional GIS-based method for extracting the geographic entities spatial relationship based on computational geometry can only perform calculations on geospatial geometric data,but cannot process geographic text data.Therefore,this paper studies the extraction method of geographic entity spatial relationship based on text semantics,which automatically extracts unstructured geographic texts in the Internet into structured geographic spatial knowledge.At present,the main problems of semantic-based spatial relationship calculation methods include polysemy of Chinese text,complex grammatical structure of geographic text,abstract entity expression,and unclear relationship semantics.In response to these problems,this paper combines neural network and natural language processing technology to study a semantic-based spatial relationship extraction method for geographic entities.The main work and innovations are as follows:(1)This paper analyzes the problem that the traditional spatial relationship extraction method based on computational geometry cannot make full use of geographic text data in the era of big data,studies the method based on text semantics,and introduces neural network technology to extract geographic entity spatial relationship from geographic text.A method of geographic entities spatial relation extraction of based on GRU-ATT is implemented.This method uses the Word2 Vec model to embed words in the text,and selects GRU to extract the features of the word vectors.On this basis,a self-attention mechanism is added to deepen the model’s ability to extract semantic features,solve the problems of complex grammatical structure of geographic text and unclear relational semantics.The experimental results show that the average F1 value of the model tested on the geographic text dataset is 70.84%,which proves the effectiveness of the model and provides a theoretical basis for subsequent research.(2)In view of the low efficiency of deep learning models,considering that some texts have only one group of entities,this paper proposes a spatial relationship extraction model based on BERT sentence semantics.The BERT sentence vector model with the best semantic representation is used to replace the Word2 Vec word embedding model which is commonly used,and the model efficiency is improved by compressing the data dimension.Since the BERT model have a multi-head self-attention mechanism and a semantic prediction task for the next sentence,the sentence semantics obtained by the model are rich in features,and the purpose of improving the accuracy and efficiency of the model is achieved at the same time.According to the comparative experiment,the average F1 value of this model is 77% when tested on geographic text data,which is 5% higher than the GRU-ATT model,and the calculation efficiency per 100 sentences of text is improved by0.67 seconds.(3)Aiming at the problem of polysemy and abstraction of geographic entity expression in Chinese text,this paper adds nature features and location features to the original BERT word vector,and fuses it to form a multi-feature word vector.The purpose of improving the representation ability of the word vector is achieved,which is beneficial to the feature extraction of the word vector.By constructing the same neural network and inputting word vectors with different feature combinations for comparative experiments,the F1 value of the model proposed in this paper is increased by 3.16%,which confirms that the fusion of multi-feature word vectors can improve model accuracy and speed up model efficiency.(4)Aiming at the problems of polysemy of Chinese text,complex grammatical structure of geographic text,abstract entity expression,and unclear relationship semantics,in the field of geography,we pioneered a geographic entity relationship extraction model based on multi-feature BERT-Bi LSTM.The model is optimized in the word embedding process by using the BERT word embedding model and integrating multiple feature vectors,the semantic expression ability is better than the One-Hot encoding and Word2 Vec encoding,and obtain the word vector with stronger representation ability,and Then input to the bidirectional long short-term memory network to extract semantics.The experimental analysis proves that the multi-feature BERT-Bi LSTM spatial relationship extraction model proposed in this study achieves an average F1 value of 78% in the geographic dataset test,which is a good improvement compared to the traditional relationship extraction model that using Word2 Vec vectorization and Attention mechanism,achieves the purpose of sovling the complex grammar structure of geographic text,polysemy of word,implicit expression of geographic entity relationship,and can adapt to the relationship extraction in geographic field.(5)Combining the previous several methods of spatial relationship extraction based on text semantics,a prototype system of geographic entity relationship query and display based on Django+Neo4j is constructed.The backend of the system will automatically crawl the text data of the Internet on a regular basis,mine the geographic entity relationship information from it,and store it in the database.The front-end of the system implements three user interaction functions: entity query,relationship query,and relationship recognition.Users can use the system to extract spatial relationships from the input text,or search for related entities or relationships on the front-end,and transmit them to the back-end by querying the Neo4 j database.The triplet pairs are displayed visually in the form of knowledge graphs.