From Text to Horn Clauses: Combining Linguistic Analysis and Machine Learning

The paper describes a system that extracts knowledge from technical English texts. Our basic assumption is that in technical texts syntax is a reliable indication of meaning. Consequently, semantic interpretation of the text starts from surface syntax. The linguistic component of the system uses a broad-coverage, domainindependent parser of English, as well as a user-assisted semantic interpreter that memorizes its experience. The resulting semantic structures are translated into Horn clauses, a representation suitable for Explanation-based Learning (EBL). An EBL engine performs symbol-level learning on representations of both the domain theory and the example provided by the linguistic part of the system. Our approach has been applied to the Canadian Individual Income Tax Guide and examples from it are used in the presentation.