Domain Adaptable Semantic Clustering in Statistical NLG
Blake Stephen Howald, Ravi Kondadadi, Frank Schilder · 2013
We present a hybrid natural language generation system that utilizes Discourse Representation Structures (DRSs) for statistically learning syntactic templates from a given domain of discourse in sentence “micro ” planning. In particular, given a training corpus of target texts, we extract semantic predicates and domain general tags from each sentence and then organize the sentences using supervised clustering to represent the “conceptual meaning ” of the corpus. The sentences, additionally tagged with domain specific information (determined separately), are reduced to templates. We use a SVM ranking model trained on a subset of the corpus to determine the optimal template during generation. The combination of the conceptual unit, a set of ranked syntactic templates, and a given set of information, constrains output selection and yields acceptable texts. Our system is evaluated with automatic, non–expert crowdsourced and expert evaluation metrics and, for generated weather, financial and biography texts, falls within acceptable ranges. Consequently, we argue that our DRS driven statistical and template–based method is robust and domain adaptable as, while content will be dictated by a target domain of discourse, significant investments in sentence planning can be minimized without sacrificing performance. 1