PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS 4.16.04 S - TREND

1. PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS4.16.04S-TREND Philip Hofmeister Neal Snider Inbal Arnon Bruno Estigarribia T. Florian Jaeger Jeanette Pettibone Ivan A. Sag

2. Superiority Effects • It has long been observed that (1) is better than (2). In fact, (2) has been labeled ungrammatical in the literature. • Who __ saw what? • *What did who see __? • Superiority Condition • In a multiple wh-question, the structurally superior wh-term is fronted/moved while the structurally inferior wh-term remains in situ • Superiority effects have been explained syntactically and pragmatically, and previous accounts have addressed both structure and function PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

3. Superiority (cont.) • Clear exceptions to the Superiority Condition, such as (3), have motivated modifications and pragmatic additions to the competence-based theory, as in D(iscourse)-linking (Pesetsky, 1987): • Which book did who read __ ? • D-linking is a separate grammatical mechanism to account for these exceptions to the Superiority Condition • This still fails to account for categorical exceptions and the gradience in acceptability PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

4. Attested Superiority ‘Violations’ • A: did you know that there are no licensing laws or sales taxes in andorra?B: I did not. What did who bring back? • http://gofree.indigo.ie/~rev/text/log/log046.html • We watch for a few more minutes as the recriminations begin: what did whom say and whati did who hear __i? But already we are bored and fearful of the ensuing calm. [Tomato is Coming, by Crispin Oduobuk] • Although nothing on this planet (or any other) can compete with the utter horror that is cilantro! Where the heck did who the heck come up with adding that gawdawful weed to otherwise civilized hote-cue-zeen? • http://www.scrappleface.com/MT/archives/001655.html • Even the starkest kind of violation is observable, e.g. • (2) *What did who see __? PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

5. Are all Superiority violations ungrammatical? NO Simple declaratives Unary wh-questions Superiority violations Ungrammatical sentences PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

6. Given the gradience and context-dependent nature of this phenomenon, the relevant constraints do not seem likely to be in the competence grammar • What other source could there be for these effects? • Processing literature offers some viable explanations • principles that apply to other cognitive processes can be invoked • overload/exhaustion of processing resources leads to unacceptability • We apply these ideas to Superiority effects PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

7. Outline • Most (if not all) of the observed superiority effects can be derived from already existing processing models • We present the results of two experiments that investigate the influence of processing factors on the acceptability of multiple wh-questions • This simplifies the grammar, providing a unified model of wh-extraction PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

8. Extending Processing Models of WH-dependencies Ease of processing wh-questions is partially determined by the interaction of processing constraints, which yield the following three preferences: I. Prefer gaps that are closer to their fillers II. Prefer fillers that have more accessible referents III. Prefer interveners that have more accessible referents Which book did who read ___? filler intervener gap PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

9. A Note on the Accessibility of WH-phrases • Wh-phrases occur without a specified, identifiable referent in mind • More accessible wh-phrases reduce the space of possible referents, on account of predetermination of candidate answer set: • which NP > (= inherently more accessible) what/who • Support from Frazier & Clifton (2002), where which-phrases are identified as better antecedents for pronouns, indicating that which-phrases mark higher accessibility. • We assume that which-NPs are easier to process than bare wh-items • Analogous to pronouns vs. definites: • The man the woman the host knew brought left vs. The man the woman you knew brought left PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

10. EXPERIMENT 1: FILLER-GAP DISTANCE

11. Method:Magnitude Estimation Acceptability judgments were collected using magnitude estimation: The judgment scale was not predetermined (participants set their own scale). The acceptability of each test item is compared with a reference sentence. Magnitude estimation was originally used in research on psychophysics but has been applied in numerous linguistic studies (see Bard et al, 1996 for more details). This does not presuppose any rigid levels of acceptability; subjects are allowed to express as many distinctions as they detect PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

12. Experimental Design • The experiments presented below were conducted on-line using the WebExp 2.1 software package (Keller et al. 1998). • Participants • all native English speakers (mean age 25.7, STDEV= 6.4). • recruited via email lists and participation did not result in any sort of payment. • Experiment 1: 41 participants • Experiment 2: 42 participants • Experiment 1: Tested Preference I (filler-gap distance) • Experiment 2: Tested Preferences II & III (accessibility of interveners and fillers) PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

13. Predicted Processing Preferences Preference 1 • Since the filler in (8) is not separated from the associated gap (8) Which man __ saw the girl in the bar on California Ave?  ‘at least as acceptable as’ (9) Which man in the bar on California Ave. did the girl see __?  (10) Which man did the girl in the bar on California Avenue see __ ? PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

14. Experiment 1: Does acceptability decrease when the distance between the filler and the gap is increased? Design – 2 x 3 EXTRACTION (subject vs. object extraction) ATTACHMENT (no-PP, wh-attachment, NP-attachment) Manipulating the distance between the wh-filler and the gap in terms of intervening discourse referents. Materials 36 items (each in 6 conditions presented below) plus 30 fillers: Subject Extraction (SE) Which man saw the girl? No attachment Which man saw [the girl in the bar on California Avenue]? NP-attachment (SE.NP) [Which man in the bar on California Avenue] saw the girl? Wh-attachment (SE.WH) Object Extraction (OE) Which man did the girl see? No attachment Which man did [the girl in the bar on California Avenue] see? NP-attachment (OE.NP) [Which man in the bar on California Avenue] did the girl see? Wh-attachment (OE.WH)

15. RESULTS Are object extractions less acceptable than subject extractions? (Probably) YES • Main effect of extraction: • significant by subjects (F1(1,35) = 4.9, p < 0.05), • near significant by items (F2(1,35) = 2.5, p = 0.12), • significant difference between means, t(40) = -2.22, p < 0.05 PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

16. Does increasing the distance reduce acceptability? YES and NO (SE.NP )Which man __ saw the girl in the bar on California Avenue ? (OE.WH) Which manin the bar on California Avenue __ did the girl see?  (OE.NP) Which man did the girl in the bar on California Ave. see __ ? (SE.WH ) Which man in the bar on California Avenue __ saw the girl? = (OE.NP) Which man did the girl in the bar on California Ave. see __ ?

17. EXPERIMENT 2: FILLER-INTERVENER ACCESSIBILITY

18. Predicted Processing Preferences Preference 2 • (11)  (12), since the filler in (11), which book, marks higher accessibility than the bare wh-word, what, in (12) : • (11) Mary wondered which book who read __  (12) Mary wondered what who read __. Preference 3 • (13)  (14), since the intervener in (13), which boy, marks higher accessibility than the bare wh-word, who, in (14): • (13) Mary wondered which book which boy read __.  (14) Mary wondered which book who read __. PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

19. Experiment 2: Does acceptability decrease when wh-fillers and interveners are less accessible? Design – 2 x 2 wh-FILLER (first wh-phrase: which NP vs. bare what) wh-INTERVENER (second wh-phrase: which NP vs. bare who) Manipulating the accessibility by choosing either the more accessible which-phrase or less accessible bare wh-phrases. Materials 20 items (all Superiority ‘violations’; each in 4 conditions presented below) plus 26 fillers: Mary wondered what who read. (BARE.BARE) Mary wondered which book who read. (WHICH.BARE) Mary wondered what which boy read. (BARE.WHICH) Mary wondered which book which boy read. (WHICH.WHICH) PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

20. RESULTS • Do less accessible wh-interveners decrease acceptability? YES • Main effect of wh-intervener accessibility (F1(1,37)= 64.5, F2(1,19)= 248.1, Ps< .001). • Do less accessible wh-fillers decrease acceptability? YES • Main effect of wh-filler accessibility (F1(1,37)= 19.2, F2(1,19)= 15.7, Ps<.001) PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

21. significant interaction between the accessibility of wh-filler & wh-intervener (F1(1,37)=9.9, F2(1,19)= 9.8, Ps<0.01) Do the accessibility of the wh-filler and the wh- intervener interact? YES • a less accessible wh-intervener (e.g. bare wh-item) always decreases acceptability, regardless of the wh-filler. • a more accessible wh-filler increased acceptability only when the wh-intervener was more accessible (“which NP”) . Mary wondered which book/what who read Mary wondered which book which boy read PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

22. Summary • Our processing model accounts for acceptability judgments of different arrangements of wh-phrases • no unpredicted effects—all but one predictions were met • Superiority, and D-linking exceptions, result from interacting processing preferences • perception of ungrammaticality results from resource limitations on memory and complexity in real-time language processing • Reinforces Arnon et al. (2005) by proving that Superiority violations are not categorically ungrammatical PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

23. Discussion: Motivations & Predictions

24. Long-distance dependency effects • Effects of distance • Gibson’s Dependency Locality Theory: “the difficulty of reactivating previous elements in the structure depends on how far back in the input they have occurred” (Gibson and Warren 2003; Grodner, Watson, and Gibson 2000) • Also, subject relatives easier than object relatives (e.g. King & Just, 1991) • Effects of accessibility of referents • “intervening elements which cause substantial integration and memory cost … are words introducing new discourse referents” (Gibson, 1998) • Similarly, our notion of accessibility for wh-NPs relies on the availability of a corresponding answer in the discourse PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

25. Similar effects of processing in other constructions? • The preferences outlined here suggest that island extractions should improve with accessibility of wh-filler (cf. Kluender 1992, 2004): • (15) Which book did they wonder whether some man had read? >= • (16) What did they wonder whether some man had read? • Again, parallel to accessibility differences between definites and indefinites: • (17) That’s the campaign that I finally thought of the aide who could spearhead. >= • (18) That’s the campaign that I finally thought of an aide who could spearhead. PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

26. Does accessibility make other predictions? • An entity’s conceptual accessibility has two basic determinants (Prat-Sala & Branigan, 2000): • derived accessibility (e.g., discourse prominence) • D-linking is an effect of increased derived accessibility • Bare wh-phrases should become more acceptable in a context • Experiment in preparation • inherent accessibility (e.g., animacy) • A processing explanation for superiority also predicts effects of the animacy of the wh-NP PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

27. Cross-linguistic evidence? • The processing account presented here also predicts that Superiority effects will occur cross-linguistically: • preferences should not be language-specific, although other language-specific factors may mask their effects • recent experiments in German (Featherston, 2004) have found Superiority effects • these were claimed not to exist by competence-based models PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

28. Conclusions A processing model of Superiority effects: • accounts for acceptability paradigms and gradience of the phenomenon • is grounded in independent, experimentally-observed factors • subsumes D-Linking under accessibility, an independently motivated interacting processing factor • maximizes explanation and minimizes grammar PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS

29. Outlook & Future Work • In ongoing work, we are using more sensitive methods, such as reading times, to isolate processing cost at integration points • experiments on the effect of context on the derived accessibility of wh-NPs • experiments on the effects of dependency distance and intervener accessibility in extraction islands • A program to recalibrate the data space for study of long-distance dependencies PROCESSING ACCOUNTS FOR SUPERIORITY EFFECTS