Geocoding Based On CRF And Spatial-relation Reasoning

Address information is closely related to human activities.In China, many departments such as statistics departments, the industry and business organizations, the public security department and so on collect and store a large amount of information containing addresses.Geocoding technology links text information containing location with spatial information, integrates spatial information and socio-economic information, and provides supporting technologies for positioning, data analysis, mapping, and visualization services. So geocoding technology plays an indispensable role in national economic construction and people’s life.Geocoding, also known as address-matching, is a process that the text address is mapped into geographic coordinates. This process constructs the relation between text position and geographic coordinates. The kernel of geocoding is address-semantic analysis and address match. So how to complete the address exact parsing by natural language processing technology and use the results to match address is the key of improving the geocoding’s accuracy. This article take the non-canonical Chinese address as the research subject, firstly the Conditional random field model is applied to address-semantic analysis, then the spatial-relation in address is applied to the address-matching. These two methods provide an available way for non-canonical Chinese geocoding, and also provide the theory and technology help for Chinese text spatial information processing and natural language spatial-querying research.The following are the main research contents:1. According to the demand and application field of geocoding, this paper firstly states the research background and significance, then analyze the research situation of geocoding from four aspects: standardization, address model, system construction and geocoding method. Here also asserts the major research content and organizational framework.2. Firstly the concept and general process of geocoding are expounded. Secondly two kernel processes are proposed in geocoding-- address parsing and address matching. Thirdly this article summarizes three methods of address parsing-- based on gazetteer(rules), based on understanding and based on the statistics, and design an address parsing method based on conditional random field. Finally it analyzes three kinds of algorithms for address matching--fuzzy address matching category, hierarchy based model algorithm and spatial reasoning based matching method.3. Building the address model based on spatial-relation, and its annotation system. On the basis of address hierarchy model and address finite automata model, this paper study address spatial-relation model and analyzes its advantage. In this paper the structure of the space-relation model and the expression of spatial relations analyzed. To find the commonly used expression patterns, a new method of analyzing address statistics based on HashTable is proposed, and this metod is used to analyze annotated address corpus.4. Address parsing based on CRF. This paper design an address-parsing method based on CRF. It focuses on the building of address corpus and address feature used to address-parsing. This method is a no-dictionary and flexible to add characteristics, so have a better performance compare to others mentioned above.5. Address-matching method based on spatial reasoning. On the basis of address parsing, this article put forward a new address-matching method based on spatial-relation reasoning, and introduces matching strategies and methods for four address expression patterns.The method makes full use of the spatial-relation information, improved matching accuracy.