Research On The Key Technology Of Named Entity Recognition And Relation Extraction In Military Field

The main object of named entities recognition on military domain is to identify the proper nouns of military document, including person, weapon, organization, location etc, which are classified into correct categories besides. It has been an important system, according to which, the commanders or command post may acquire military intelligence quickly and make milit ary policy in time and accurately. Recently the named entities recognition on military domain has been an important research aiming at improving military automation and intelligence. Based on this technology, the military document in form of unstructured free text are analyzed and processed, by means of the technologies of language processing, including automatic segmentation, classification tagging technology and information extraction technology et, then the information to be involved with command and control of armed forces are extracted and are changed into structured data that can be identified by computer rapidly and accurately, eventually the system generate the strategic deployment.In our research, we studied on several important technologies theoretically and technologically, which was used in named entities recognition on military domain. First, we established three dictionary matching, the military text as the data source, and put the military documents through word segmentation, using the forward maximum matching method. Second, design category labels. Every word segmentation unit was labeled with a category label, which was used to identify and extract in the category. Third, our study put conditional random fields through iterative learning control, using Tri-Training algorithm. After key words data was identified and extracted, the preliminary identified data was corrected using the dictionary based correction method, and the extracted data was written in the structured documents, in order to accomplish conversion of the text into structured data. In our study, through the comparison of the four groups of experimental data, it was showed that more compound words and nested words can be identified accurately by using the dictionary based correctio n method, so that the correct rate of recognition and the recall rate was improved greatly. This method may accomplish named entities recognition on military domain excellently. When the operational documents and network military text data were tested, the highest F- value was 92.40%, which was similar to the level of named entities recognition on general domain.