Query analysis using context-based information extraction on navigation domain

Information extraction is an appropriate methodology to process queries in purposeful dialogs such as navigation and question answering. We apply information extraction to find meaningful instances from queries on navigation domain and to fill them into predefined slots. To deal with various ambiguities and flexible utterances of queries, we introduce three-leveled robust rules; instance extraction, filtering, and ranking rules. The extraction rules select instance candidates by matching the input query with lexico-semantic patterns. Using the context consisting of lexico-semantic patterns and slot names, the filtering rules remove improper ones among the candidates, and the ranking rules endow them with positive scores. We choose navigation domain as a test environment to experiment the adaptability of information extraction. Our experimental result shows the understandability against user´s queries increases up to 97% even though speech recognition has only the accuracy of 75.2% in sentence level.