Inducing lexical entries for an incremental semantic grammar

We introduce a method for data-driven learning of lexical entries in an inherently incremental semantic grammar formalism, Dynamic Syntax (DS). Lexical actions in DS are constrained procedures for the incremental projection of compositional semantic structure. Here, we show how these can be induced directly from sentences paired with their complete propositional semantic structures. Checking induced entries over an artificial dataset generated using a known grammar demonstrates that the method learns lexical entries compatible with those defined by linguists, with different versions of the DS framework induced by varying only general tree manipulation rules. This is achieved without requiring annotation at the level of individual words, via a method compatible with work on linguistic change and routinisation.