A Dynamic Approach for Automatic Error Detection in Generation Grammars

In any real–world application scenario, natural language generation (NLG) systems have to employ grammars consisting of tremendous amounts of rules. Detecting and fixing errors in such grammars is therefore a highly tedious task. In this work we present a data mining algorithm which deduces incorrect grammar rules by abductive reasoning out of positive and negative training examples. More specifically, the constituency trees belonging to successful generation processes and the incomplete trees of failed ones are analyzed. From this a quality score is derived for each grammar rule by analyzing the occurrences of the rules in the trees and by spotting the exact error locations in the incomplete trees. In prior work on automatic error detection v.d.Brück et al. [5] proposed a static error detection algorithm for generation grammars. The approach of Cussens et al. creates missing grammar rules for parsing using abduction [1]. Zeller introduced a dynamic approach in the related area of detecting errors in computer programs [6].