A Best-Fit Approach to Productive Omission of Arguments

and Fillmore 1999) provide a framework for describing grammaticality in terms of both form and meaning constraints. On the meaning side, cognitive linguists have contributed many great insights into the conceptual structures necessary for understanding language, but challenges remain in developing a precise, formal representation of construction grammar that supports both detailed linguistic analysis and computational use. One such challenge arises from pro-drop languages, which allow productive argument omission where arguments can be omitted without any markedness in use. Mandarin Chinese is one such example where both the subject and the object can be freely omitted. This phenomenon is commonplace in conversations, but has also been shown in written text (Li 2004; Yeh and Chen 2004). In actual use, different arguments of a construction tend to be omitted at different rates. Additionally, our preliminary data in Mandarin suggests that even semantically related arguments are omitted differentially in different constructions. This data is difficult to properly account for, both linguistically and compu-tationally, by a general principle of omission that is applied uniformly to all constructions. For a theory of grammar to accurately account for such data, the grammar must incorporate construction-specific rates of argument omission. Unfortunately, specifying such parameters in most theories of construction grammar requires enumeration of all possible argument combinations as separate constructions within the grammar. This leads to not only an explosion of constructions but also a loss of generality. However, situating the grammar within a best-fit processing model allows us to avoid such pitfalls. Intuitively, a best-fit processing model finds the interpretation of an utterance that fits the utterance best. More precisely, the model incrementally builds up a set of competing interpretations for an utterance, attending only to those interpretations that are most likely given the probabilistic syntactic, contextual, and