Quantitative Evaluation of Explanation-Based Learning as an Optimisation Tool for a Large-Scale Natural Language System

This paper describes the application of explanation-based learning, a machine learning technique, to the SRI Core Language Engine, a large scale general purpose natural language analysis system. The idea is to bypass normal morphological, syntactic and (partly) semantic processing, for most input sentences, instead using a set of learned rules. Explanation-based learning is used to extract the learned rules automatically from sample sentences submitted by a user and thus tune the system for that particular user. By indexing the learned rules efficiently, it is possible to achieve dramatic speedups. Performance measurements were carried out using a training set of 1500 sentences and a separate test set of 100 sentences, all from the ATIS corpus. A set of 680 learned rules was derived from the training set. These rules covered 90 percent of the test sentences and reduced the total processing time to a third. An overall speed-up of 50 percent was accomplished using a set of only 250 learned rules.