Statistically-guided controlled language authoring

Abstract

This study presents a series of experiments using contextual next word prediction to aid controlled language authoring. The goal is to assess the capabilities of n-gram language modelling to improve the generation of controlled language in restricted domains with minimal supervision. We evaluate how different dimensions of language model design can impact prediction and textual coherence. In particular, evaluations of suggestion ranking, perplexity gain and language model combination are presented. We show that word prediction can provide adequate suggestions which could offer an alternative to costly manual configuration of rules in controlled language applications.