A Probabilistic Approach to Error Detection&Correction for Tree-Mapping Grammars
Abstract Rule-based natural language generation denotes the process of converting a semantic input structure into a surface representation by means of a grammar. In the following, we assume that this grammar is handcrafted and not automatically created for instance by a deep neural network. Such a grammar might comprise of a large set of rules. A single error in these rules can already have a large impact on the quality of the generated sentences, potentially causing even a complete failure of the entire generation process. Searching for errors in these rules can be quite tedious and time-consuming due to potentially complex and recursive dependencies. This work proposes a statistical approach to recognizing errors and providing suggestions for correcting certain kinds of errors by cross-checking the grammar with the semantic input structure. The basic assumption is the correctness of the latter, which is usually a valid hypothesis due to the fact that these input structures are often automatically created. Our evaluation reveals that in many cases an automatic error detection and correction is indeed possible.