More for less: learning a wide covering grammar from a small training set

This paper describes a grammar learning system which combines model-based and data-driven learning within a single framework. Results from learning grammars with the Spoken English Corpus (SEC) suggest that a combined model-based and data-driven learner can acquire a wide coverage grammar from only a small training corpus. In this paper, we present some results of our grammar learning system. We show that using uniication-based grammars, with a hybrid learning system allows a rapid rate of convergence upon a test corpus with only a modest amount of training material. In contrast to other researchers (for example (BMMS92; GLS87; Bak79; LY90; VB87)), we try to learn competence grammars and not performance grammars. We also try to learn grammars that assign linguistically plausible parses to sentences. Learning competence grammars that assign plausible parses is achieved by combining model-based and data-driven learning within a single framework (OB93b; OB93a; Os-bng). Model-based (deductive) methods are sound (MKKC86) (assuming that the model is consistent), but suuer from incompleteness, whilst data-driven (inductive) methods are unsound (they cannot guarantee that natural languages can be learnt (Gol67)), but complete. Note that`completeness' here means that the learner is always in a position to make a decision. We let both of the learning styles compensate for each other's weaknesses. A recent result showed that the combined use of induction and deduction produced a grammar that assigned quantitatively more plausible parses to sentences taken from the Spoken English Corpus (SEC) (LG91) than is the case when using either learning style in isolation (OB94). The system is implemented to make use of the Grammar Development Environment (GDE) (CGBB88) and it augments the GDE with 3300 lines of Common Lisp. The structure of this paper is as follows. Section 2 gives an overview of the combined model-based and data-driven learner. Section 3 then describes the method used to generate the results, which are then presented in section 4. Section 5 discusses these results and points the way forward. 2: System overview Architecture We assume that the system has some initial grammar fragment, G, from the outset. Presented with an input string, W, an attempt is made to parse W using G. If this fails, the learning system is invoked. Learning takes place through the inter-leaved operation of a parse completion process and a parse rejection process. In the parse completion process, the learning system tries to generate rules that, had they been members of G, …