Layers of Interpretation: On Grammar and Compositionality

With the recent resurgence of interest in semantic annotation of corpora for improved semantic parsing, we observe a tendency which we view as ill-advised, to conflate sentence meaning and speaker meaning into a single mapping, whether done by annotators or by a parser. We argue instead for the more traditional hypothesis that sentence meaning, but not speaker meaning, is compositional, and accordingly that NLP systems would benefit from reusable, automatically derivable, taskindependent semantic representations which target sentence meaning, in order to capture exactly the information in the linguistic signal itself. We further argue that compositional construction of such sentence meaning representations affords better consistency, more comprehensiveness, greater scalability, and less duplication of effort for each new NLP application. For concreteness, we describe one well-tested grammar-based method for producing sentence meaning representations which is efficient for annotators, and which exhibits many of the above benefits. We then report on a small inter-annotator agreement study to quantify the consistency of semantic representations produced via this grammar-based method.