Explanation-based Learning Helps Acquire Knowledge from Natural Language Texts

Existing systems to acquire knowledge from expository texts do not perform any learning beyond interpreting the contents of the text. The opportunity to learn from examples included in texts is not exploited. This is a needless limitation because examples in texts usually show the reader how to integrate the declarative part of the text into an operational concept or procedure. Explanation-based Learning (EBL) seems to fill this gap as it explains the example within the domain theory, generalizes the explanation and operationalizes the concept definition by compiling necessary knowledge from the domain theory into the definition. In this paper, we study the synergistic combination of automatic text analysis and EBL. EBL is used realistically, where the domain theory and the training examples are obtained from a specification or a regulation by a text analysis program, rather than being given a priori. We present a prototype system which demonstrates the potential of this approach. The paper includes a detailed example using the Canadian Income Tax Guide.