Place-related question answering: From questions to relevant answers

Abstract

In everyday communications, people talk about space by referring to places. While the common sense notion of place is understandable to humans, formalising place in a computational model remains a challenging issue. The strong context dependency, diverse metaphorical uses, indeterminacy of boundaries, and vernacular reference use are major challenges in making place knowledge digestible for computers. This research aims to utilise domain knowledge to study place-related questions and their corresponding answers, and to develop models and methods to answer the questions.

In the context of place-related question answering, this study investigates what people expect from computers to understand about places, and how these place-related questions are answered in human-generated responses. First, a place model is designed for the question answering purpose using the collective domain knowledge extracted from literature. Later, the model is used to characterize the platial information in place-related questions and their human-generated answers. In the next step, the natural language questions are translated to GeoSPARQL queries to enable the spatial analysis for answering place-related questions. Finally, templates for answering where-questions are proposed to generate relevant responses similar to human-generated answers.

The results of this study show that domain knowledge can be used to improve current methods of place-related question answering. Using domain knowledge, an encoding method is devised that can characterise large question answering corpora with minimal supervision. The encoding results are used to identify descriptive patterns inside the questions and answers.

In the next step, a novel approach is designed using domain knowledge and object-based conceptualization of place to translate natural language questions to GeoSPARQL queries. The novelty of the approach is mainly to (1) use domain knowledge and avoid reinventing new terms, and (2) utilise FOL statements as the intermediate representation which can be later translated not only to GeoSPARQL but any other formal query languages with minimal efforts. The method is tested using the Geospatial Gold Standard dataset, and the results show significant improvements in extracting information and translating questions to queries in comparison to the state-of-the-art approaches.

Finally, the relevance of answers to where-questions is investigated using templates of generic information (i.e., type, scale and prominence). The results show that generic representations can be used to characterise answers in a few frequent patterns and also to study relevance of answers to the questions. Moreover, the extracted knowledge can be captured using sequence prediction methods in a machine digestible manner. The results of this study can be used to test the relevance of machine-generated responses or to generate automatic responses similar to human-generated answers.

Overall, this thesis contributes to the domain of geographic question answering with a focus on geographic places. The results of this study can be used in question answering systems to analyse and classify the questions, generate queries and formulate relevant responses. The results of this study show the importance of domain knowledge in improving the performance of existing question answering systems, and also provide useful insights about human answering behaviour.