Semantics-Based Reference Resolution in Technical Text Processing: An Exploration of Using the WordNet Database in the Computerized Comprehensibility System.

摘要

The Computerized Comprehensibility System (CCS) provides an automated copy editing function, generating a 'mark-up' of a draft of a technical document by simulating the simpler comprehension processes of a human reader, and then criticizing the text when these simple processes cannot successfully comprehend the material. A key CCS function is criticizing the coherence of the material by tracking which objects are mentioned in the passage. A common comprehensibility problem is that the text mentions a new object using the syntactic structures appropriate for an already-known object. If the reader must make an inference that presence of the new object is implied by earlier-mentioned object, the result is a potential break in the coherence of the text. CCS criticizes all such coherence breaks. However, many such inferences are actually easy for most readers, since only general knowledge is required to make the inference, rather than specialized knowledge about the domain. If so, then the CCS criticism of a coherence break is a false alarm. This report describes exploratory work with an augmented form of CCS, in which the WordNet database is used as a source of general knowledge to allow CCS to make the same kind of general knowledge inferences that human readers do to overcome coherence breaks. Training, Learning, Machine Learning, Explanation-based Learning.