Using decision trees to infer semantic functions of attribute grammars

Abstract

In this paper we present a learning method called LAG (Learning of Attribute Grammar) which infers semantic functions for simple classes of attribute grammars by means of examples and background knowledge. This method is an improvement on the AGLEARN approach as it generates the training examples on its own via the effective use of background knowledge. The background knowledge is given in the form of attribute grammars. In addition, the LAG method employs the decision tree learner C4.5 during the learning process. Treating the specification of an attribute grammar as a learning task gives rise to the application of attribute grammars to new sorts of problems such as the Part-of-Speech (PoS) tagging of Hungarian sentences. Here we inferred context rules for selecting the correct annotations for ambiguous words with the help of a background attribute grammar. This attribute grammar detects structural correspondences of the sentences. The rules induced this way were found to be more precise than those rules learned without this information.