Accurately recognizing grammatical dependencies in an asyntactic memory
Matthew Wagers (New York University & University of California, Santa Cruz)
November 18, 2008 (Tuesday)
6:30 PM - 8:00 PM; Room 7102, CUNY Graduate Center

The mental operations by which wh-dependencies are parsed have been studied intensively in the past 30 years (Wanner & Maratsos, 1978; Fodor, 1978; Stowe, 1986; Frazier, 1987, etc.). The conclusions from this literature are notably convergent: wh- dependencies are parsed (1) actively and (2) accurately. (1) The comprehender projects hypotheses about the complete wh-chain in advance of local evidence for the position of an empty category. (2) The comprehender almost always seems to generate candidate hypotheses that are grammatical (see Phillips & Wagers, 2007). A comprehender who achieves (1) and (2) must have a reliable representation of syntactic context that can be used to make decisions about incoming input and to recover the information in previously-encoded constituents (cf. Berwick & Weinberg, 1984). However, I will argue that an emerging 'new synthesis' on human short-term memory (McElree, 2006, Jonides, et al., 2008) presents a challenge for meeting these requirements. Firstly, the scope of the active representational workspace is claimed to be severely limited, which biases decision-making with extremely local information. Secondly, the means of retrieving constituent representations is claimed to be indexed to content and not syntactic order. Decision-making is thus biased toward the inherent properties of particular constituent encodings, rather than relations that hold between encodings like c-command or dominance. Some evidence for grammatical fallibility in the licensing or interpretation of (non-A') dependencies supports this view (e.g., Meng & Bader, 2000, Van Dyke & Lewis, 2003, Vasishth, et al., 2005). It is however also qualified by observations of grammatical fidelity in similar structures (e.g., Sturt, 2003, Wagers, et al., 2008, Xiang et al., 2008). In this talk I hope to show that the apparent clash between the memory architecture and the nature of syntactic representation, while real, is not insurmountable. Moreover, it provides a productive research framework precisely because it so strongly constrains our hypothesis about how information is encoded and retrieved. In this context, I will discuss five real-time studies on wh-dependency comprehension that address: (a) what kinds of information can be maintained over the length of the dependency; (b) whether a unique dependency head can be targeted for retrieval by the parser, despite the presence of highly similar (wh)-constituents in the same sentence. I conclude that the parser's expectations about future input support the strategic encoding of the current input, in a way that enables grammatically-faithful dependency licensing.