SYNTACTIC ADAPTATION 1. Rapid Expectation Adaptation During Syntactic Comprehension. Alex B. Fine * T. Florian Jaeger. Thomas A. Farmer.

Transcription

1 SYNTACTIC ADAPTATION 1 Running head: SYNTACTIC ADAPTATION Rapid Expectation Adaptation During Syntactic Comprehension Alex B. Fine * T. Florian Jaeger Thomas A. Farmer Ting Qian University of Rochester Department of Brain and Cognitive Sciences * Please address correspondence to: Alex B. Fine Brain & Cognitive Sciences Meliora Hall, Box University of Rochester Rochester, NY USA +1 (585) afine@bcs.rochester.edu

2 SYNTACTIC ADAPTATION 2 Abstract We report two experiments testing the hypothesis that humans rapidly adapt to the statistics of novel linguistic environments during language comprehension, focusing on syntactic comprehension. Two reading experiments provide evidence for rapid adaptation in syntactic comprehension. We conclude that humans are able to quickly adjust their subjective representation of the statistics of the environment. The results extend the literature on syntactic priming, and suggest a link between adaptation in syntactic comprehension and adaptation in lower-level, perceptual domains, where most previous work on adaptation has been concentrated. Keywords: adaptation, syntactic comprehension, implicit learning, syntactic priming

3 SYNTACTIC ADAPTATION 3 Rapid Expectation Adaptation During Syntactic Comprehension Introduction Language use is variable, in the sense that each speaker of a language enters any given communicative situation with a unique set of life experiences that shape, most basically, the way that person says things. For example, many people can relate to the experience of participating in a conversation in which both participants are native speakers of the language in use, but, based on differences in life experience, speak with different accents or in different regional dialects (say, when someone from the Washington, D. C., area speaks English with someone from a small town in northern Scotland). Or, consider situations in which it is necessary to effectively converse with someone who is not a native speaker of the language in use. Although situations such as these can often produce humorous instances of misunderstanding as when Parker Pens famously and inadvertently claimed in a Mexican ad that its pens would not leak in your pocket and impregnate you, intending to use the Spanish verb embarazar ( to impregnate ) to mean embarrass communication is typically successful. That is, with sufficient time, the systems responsible for real-time language comprehension seem to be able to accommodate considerable multi-dimensional variability in the linguistic signal. Indeed, as we discuss in below, variability in the linguistic signal is the norm, not the exception. In the literature on speech perception, this is known as the infamous lack of invariance (Liberman, Cooper, Shankweiler, & Studdert-Kennedy, 1967): Simply put, realizations of the same sound category (phoneme) are realized differently by different

4 SYNTACTIC ADAPTATION 4 speakers of the same language (Labov, 1966), as well as speakers that for all relevant purposes would be classified as speaking the same dialect and sociolect (Allen, Miller, & DeSteno, 2003). Even productions of the same sound by the same speaker differ over time (Bauer, 1985; Harrington, Palethorpe, & Watson, 2000; Yaeger-Dror, 1994) and they can differ at a given time depending on both linguistic context (such as the surrounding sounds; Nolan, Holst, & Kuehnert, 1996; Ohman, 1966; Whalen, 1990) and social context (such as the register or speech style;(bell, 1984). In short, although less obvious than the kinds of variability mentioned at the beginning of this introduction, listeners face an enormous amount of between- and within-speaker variability. The question of how listeners overcome the challenges for speech perception that result from this variability has been one of the central questions in phonetics and psycholinguistics research over at least the last five decades (e.g., Bradlow & Bent, 2008; Clarke & Garrett, 2004; Kraljic & Samuel, 2005, 2006, 2007; Kraljic, Samuel, & Brennan, 2008; Liberman et al., 1967; Norris, McQueen, & Cutler, 2003; Vroomen, van Linden, de Gelder, & Bertelson, 2007). The consensus that emerges from this work is that the systems underlying speech perception address the problem of invariance by (a) adapting to changes in the realization of sounds (e.g., Clayards, Tanenhaus, Aslin, & Jacobs, 2008; Eisner & McQueen, 2005; Kim, Horton, & Bradlow, 2011; Kraljic & Samuel, 2007; Pardo, 2006) and (b) storing rich representations associated with specific speakers and situations (e.g, Goldinger, 1996, 1998; Johnson, 1997; Nosofsky, 1986; Pierrehumbert, 2002; Tenpenny, 1995). We refer to the former as phonetic adaptation. While phonetic adaptation has been the subject of a long and rich research tradition, comparatively little is known about how listeners deal with linguistic variability beyond

5 SYNTACTIC ADAPTATION 5 the level of phonetic processing. Here, we focus on the question of whether the systems underlying syntactic processing exhibit the same ability to adapt to situation- and speakerspecific variability in the structural realization of sentences, as has been observed for speech perception. Here we use the term adapt (and adaptation ) to refer to the processes that enable listeners to adjust their linguistic expectations based on previous experience to a specific speaker or situation. In this context, it is important to note that the same or similar concept that we refer to as adaptation has been discussed under a variety of different labels often without further definition, including entrainment (Gravano et al., 2011; Levitan & Hirschberg, 2011; Levitan, Gravano, & Hirschberg, 2011), accommodation (Giles, Coupland, & Coupland, 1991; M. Kim et al., 2011; Pardo, 2006), and alignment (e.g., Pickering & Garrod, 2004). We use the term adaptation as it highlights parallels to research beyond speech perception, including visual perception (e.g., Blakemore & Campbell, 1969; Leopold, O Toole, Vetter, & Blanz, 2001 for behavioral work, and Barlow & Hill, 1963; Movshon & Lennie, 1979 for work at the cellular level) and motor planning (e.g., Koerding, Tenenbaum, & Shadmehr, 2007; McLaughlin, 1967; Shadmehr & Mussa-Ivaldi, 1994; Wallman & Fuchs, 1998). We consider this parallel particularly revealing for reasons outlined in the general discussion We use the term accommodation to refer to what is presumably the purpose of adaptation, namely the ability to maintain robust and perhaps efficient language understanding in the face of extensive crosssituational variability. Two more clarifications about the term adaptation are in order before we can proceed. First, we leave open for now whether adaptation involves changes in linguistic

6 SYNTACTIC ADAPTATION 6 representations or rather changes in how these representations are accessed or employed during sentence understanding. Second, there is an ongoing debate centering on the question of whether linguistic adaptation is due to short-term activation changes (sometimes called priming, e.g., Branigan, Pickering, & Cleland, 1999; Malhotra, 2009) or rather due to implicit learning (e.g., (Bock & Griffin, 2000; Chang, Dell, & Bock, 2006; Kraljic et al., 2008). We use the term adaptation in a mechanism-neutral sense to refer to adaptive behavior regardless of whether it is due to learning or other mechanisms. However, we return to this question in the general discussion. With these clarifications in mind, we define syntactic adaptation as the processes that allow listeners to adjust their syntactic expectations to specific speakers or situations. The first goal of the current paper is to explore whether there is syntactic adaptation and to what extent it resembles what is known about phonetic adaptation. Our second goal is to integrate the various lines of research that are relevant to linguistic adaptation in general and to situate this research in the broader context of perceptual and cognitive adaptation beyond the linguistic domain. Despite striking similarities across domains, research on adaptation has often been pursued in isolated lines of research. Because previous work on adaptation has been concentrated in lower-level perceptual domains (speech perception, depth perception, motor planning, etc.), finding such a resemblance would, in turn, contribute towards establishing adaptation as a general principle of human perception and cognition. We begin by summarizing several important generalizations that emerge from the literature on phonetic perception relevant to adaptation. This will set the stage for a discussion of syntactic adaptation: what is it, when we expect to observe it, and how we

7 SYNTACTIC ADAPTATION 7 plan to investigate it. The Lack of Invariance in Speech We started this paper with an example of cross-speaker variability that is due to radically different language backgrounds (speakers of different dialects or different languages). While examples like this illustrate the lack of invariance, they are comparatively rare in the lives of most speakers. The much more common, although less obvious, challenge to speech perception comes from pronunciation variability between speakers of the same language (Labov, 1966), dialect (e.g., Allen et al., 2003; Eckert, 2000), or even the same speaker in different situations (Antoniou, Best, Tyler, & Kroos, 2010; Nearey & Rochet, 1994). Such variability involves subtle changes in how a phonemic category boundary is mapped onto an acoustic or phonetic feature that distinguishes between these phonemes (e.g., voice onset timing, or VOT, as a cue to voicing, which contributes to the perceived phonological contrast between the English phonemes /p/ and /b/). This variability could also involve changes in how different features are weighted together in creating a phonological contrast (Clayards, 2009; Hazan & Baker, 2011). For instance, Figure 1, taken from Allen et al. (2003), illustrates variability along the dimensions of both word duration and VOT for the word pace for two speakers of English. This figure underscores just how noisy and non-deterministic the relationship is between perceptual input available to the listener (word duration, VOT) and the linguistic categories corresponding to that perceptual input. This variability is often belied by our subjective impression of invariance across and within speakers XXXXX Insert Figure 1 XXXXX------

8 SYNTACTIC ADAPTATION 8 In light of the pervasive lack of invariance in the speech signal, it should come as little surprise that researchers in linguistics and cognitive science have for a long time been fascinated by the question of how we can understand each other at all. Part of the answer that has emerged over the last few decades is that listeners can rapidly adapt to speech signals that deviate from their previous experience. For example, as little as one minute of exposure can undo the initial disadvantage associated with the processing of non-native accented speech (Clarke & Garrett, 2004), and this disadvantage is substantially reduced after as few as 4 sentence-length utterances (Bradlow & Bent, 2008; Clarke & Garrett, 2004; Weil, 2001). Similarly, rapid adaptation has been observed to dialectal variants of the listener s own native language (e.g., Best, McRoberts, & Goodell, 2001), as well as to distorted speech (Remez, Rubin, Berns, Pardo, & Lang, 1994; Remez, Rubin, Pisoni, & Carrell, 1981; Rosen, Faulkner, & Wilkinson, 1999; Shannon, Zeng, Kamath, Wygonski, & Ekelid, 1995) and synthetic speech (Golomb, Peelle, & Wingfield, 2007; Greenspan, Nusbaum, & Pisoni, 1988; Pallier, Sebastian-Gallés, Dupoux, Christophe, & Mehler, 1998; Sebastián-Gallés, Dupoux, Costa, & Mehler, 2000; Wingfield, Peelle, & Grossman, 2003). While an increasingly large body of work explores the properties of adaptation in phonetic perception, much less is known about adaptation at higher levels of language processing, such as syntactic processing. This is despite the fact that, as we show below, the lack of invariance in the linguistic signal is not restricted to the mapping between acoustics and phonology. Speakers differ in terms of their preference for lexical items (e.g., couch vs. sofa) and syntactic structures (e.g., the choice between Hand that man a

9 SYNTACTIC ADAPTATION 9 banana and Hand a banana to that man). Just like variability at the phonetic level can affect the speed and accuracy of word recognition (Gaskell & Marslen-Wilson, 1996), variability in syntax is expected to affect syntactic processing (for evidence, see (Ferreira, Kleinman, Kraljic, & Siu, in press; Kaschak & Glenberg, 2004). This raises the question as to whether the systems involved in real-time language processing are set up to (rapidly) overcome the challenges that result from syntactic variability. This is the question we begin to address in this paper. We ask whether the systems underlying the incremental understanding of sentences exhibit the same flexibility observed for phonetic perception: can listeners adapt to speaker- or situation-specific variation at the syntactic level? If so, is such adaptation observed at similarly short timescales? If so, this might point to adaptation as a general property of the language processing system, rather than being limited to low-level perception (cf. recent work on phonetic adaptation, which has begun to ask whether such adaptation is limited to perception or involves at least reference to phonological and lexical information; Bradlow & Bent, 2008; Kraljic & Samuel, 2006). Such a finding would in particular be expected by accounts that assume life-long learning over everything we process (Chang et al., 2006; Elman, 1990; MacDonald & Christiansen, 2002; see also Botvinick & Plaut, 2004). For example, in the connectionist model of syntactic acquisition proposed by Chang and colleagues (Chang et al., 2006), they propose that the prediction error resulting from processing a sentence (the degree to which the sequence of words in a sentence deviates from what was expected) is used to adjust the weights of the model for successive sentences. This implicit learning is explicitly assumed to continue throughout adult life ( We assume that the brain s learning and processing mechanisms function throughout life

10 SYNTACTIC ADAPTATION 10 (learning-as-processing assumption), Chang et al., 2006, p. 240). Life-long learning is also assumed by exemplar-based models (Bod, 2006; Johnson, 1997; Pierrehumbert, 2002). To the best of our knowledge, this prediction has, however, not been put to as direct a test as we conduct below. Experience-based Language Processing The starting point for our investigation is the observation that listeners seem to take advantage of statistical contingencies in the input during sentence understanding (e.g., MacDonald, Pearlmutter, & Seidenberg, 1994). For example, there is broad agreement that listeners are able to anticipate upcoming words and structures based on cues in the preceding parts of an utterance (Boston, Hale, Vasishth, & Kliegl, 2011; Demberg & Keller, 2008; Frank & Bod, 2011; Kamide, Altmann, & Haywood, 2003; McDonald & Shillcock, 2003; Staub & Clifton, C., 2006). Statistical contingencies also play a crucial role in the processing of temporarily ambiguous material in so-called garden path sentences (e.g., Garnsey, Pearlmutter, Myers, & Lotocky, 1997; MacDonald et al., 1994; McRae, Spivey-Knowlton, & Tanenhaus, 1998; Tanenhaus, Spivey-Knowlton, Eberhard, & Sedivy, 1995; Trueswell, Tanenhaus, & Garnsey, 1994; Trueswell, Tanenhaus, & Kello, 1993). Consider, for instance, sentences such as (1), for which the syntactic assignment of the noun phrase (NP) Agent Cooper is temporarily ambiguous. It is interpretable both as the direct object (DO) of the matrix verb believed, as in (1a), and as the subject of a sentence complement (SC) to the matrix verb, as in (1b). (1) Audrey believed Agent Cooper...

11 SYNTACTIC ADAPTATION 11 a.... because he had never lied to her. b....was in love with her. Compared to unambiguous sentences like (2), the disambiguating region in (1b) was, in this example results in processing difficulty (as reflected in, for example, higher reading times; see Frazier, 1987). We will refer to the difference in reading times, at the point of disambiguation, between unambiguous and ambiguous sentences as the ambiguity effect. (2) Audrey believed that Agent Cooper was in love with her. The size of the ambiguity effect in garden-path sentences such as (1)-(2) is known to correlate with how expected the intended interpretation is, based on a variety of cues preceding the disambiguating region (e.g, Garnsey et al., 1997; MacDonald et al., 1994; McRae et al., 1998; Tanenhaus et al., 1995; Trueswell et al., 1994, 1993). For instance, the ambiguity effect in sentences like (1) and (2) is affected by the so-called subcategorization bias of the matrix verb (believed in (1) and (2)): verbs that are more likely to take a sentence complement are associated with a smaller ambiguity effect on the disambiguating region (Trueswell et al., 1993). Similarly, other cues, such as the suitability of the ambiguous NP as a subject of a sentence complement (rather than the direct object to the matrix verb), have been shown to further reduce the ambiguity effect (Garnsey et al., 1997). Findings like these have constituted one of the central motivations for so-called

12 SYNTACTIC ADAPTATION 12 cue-based, competition, constraint satisfaction, and in the most general terms experience-based accounts of language processing (Bates & MacWhinney, 1987; Cuetos, Mitchell, & Corley, 1996; Hale, 2001; Jurafsky, 1996; Levy, 2008; MacDonald et al., 1994; Mitchell, Cuetos, Corley, & Brysbaert, 1995; Narayanan & Jurafsky, 2002; Reali & Christiansen, 2007; Tabor, Juliano, & Tanenhaus, 1997; Trueswell et al., 1994). All of these accounts share the assumption that listeners previous language experience shapes their language processing system (Elman, 1990; MacDonald, 1999). That is, listeners are assumed to process language based on their subjective beliefs about the language based on their individual previous experience. This is a crucial assumption that so far has mostly remained untested. The works cited above all employed estimates of syntactic expectations based on norming studies that averaged across many participants or based on language corpora (which also implicitly averages across speakers). While tests based on such estimates of average experience constitute an important step in establishing experience-based accounts, even stronger evidence for experience-based accounts would come from finding that individual differences in language experience are reflected in expectation-based processing difficulty. Indeed, a handful of recent studies support this prediction of experience-based accounts (Farmer, Misyak, Fine, & Christiansen, under review; Fine & Jaeger, under review; MacDonald & Christiansen, 2002; Wells, Christiansen, Race, Acheson, & MacDonald, 2009). All of these studies present evidence that sizeable differences in individual experiences affect language processing. They do not, however, provide evidence that recent experience immediately affects syntactic comprehension. If the language comprehension system exploits statistical information in the linguistic signal that has been

13 SYNTACTIC ADAPTATION 13 accumulated over the course of the comprehender s life, we might expect comprehenders to engage in a process of continuous adaptation, updating their beliefs about the ambient language in accordance with their experience with that language. It is possible, of course, that language experience, while continuously monitored, is integrated only with a delay (e.g., during sleep, as some researchers have hypothesized for language and for other cognitive domains, cf. Davis & Gaskell, 2009; Fenn, Nusbaum, & Margoliash, 2003; Karni, A. & Sagi, 1993; C. Smith & Rose, 1997). On the other hand, learning could be truly continuous, so that its effects are evident after very limited exposure. The latter is arguably expected if adaptation serves to accommodate speakerspecific production preferences: adapting to a specific speaker only after a night s sleep will often be of little use. If this is true, then the experience-based effects on language processing based on estimates from corpora or norming studies which presumably reflect subjects experience on average with a variety of linguistic events should be modulated by the statistical properties of the immediate linguistic situation. This is one of the predictions tested by our experiments. Our experiments also test a more specific assumption shared by many experiencebased accounts that has often been left implicit: the assumption that listeners take advantage of statistical regularities in order to process language efficiently (Hale, 2001; Jurafsky, 1996; Levy, 2008; Seidenberg & MacDonald, 1999). This idea is perhaps made most explicit in recent work that approaches the problem of language comprehension from the perspective of optimal resource allocation or rational inference under uncertainty (Levy, 2011; Levy, Bicknell, Slattery, & Rayner, 2009; Norris et al., 2003; N. J. Smith & Levy, 2008). This work is built on the assumption that listeners have access to implicit

14 SYNTACTIC ADAPTATION 14 knowledge of the statistical relationship between cues and linguistic structures. Under the (implicit) assumption that the expectations of comprehenders are sufficiently closely aligned with the production preferences of their interlocutors, this work shows that reliance on statistical regularities would theoretically enable listeners to optimally allocate processing resources (N. J. Smith & Levy, 2008), minimize processing costs (Levy, 2008) and optimally cope with noisy perceptual input (Levy, 2011; Levy et al., 2009; Norris & McQueen, 2008). What this line of work has not discussed is that reliance on statistical regularities might actually have the opposite effect if listeners expectations deviate strongly from their interlocutors production preferences. The expectations of interlocutors are bound to be closely aligned with the producers preferences only if at least one of the following two scenarios holds: (1) syntactic preferences are sufficiently stable between speakers, and listeners acquire knowledge about these preferences early in their lives or (2) listeners can (rapidly) adapt to differences between speakers. In the following section, we argue that (1) is unlikely to be the case. We summarize evidence that syntactic production exhibits considerable variability, just as illustrated above for phonetic production. This leads us to entertain the possibility of rapid syntactic adaptation (i.e., scenario 2). We discuss existing evidence that can be interpreted as syntactic adaptation. Then we present two self-paced reading experiments that assess whether and, if so, to what extent readers adapt to situation-specific changes in the statistical regularities underlying lexical cues to syntactic structure. Specifically, we investigate whether readers can overcome the processing disadvantage associated with garden path sentences if they receive sufficient evidence that these structures are frequent in the context of the experiment and should therefore be expected. If so, this would

15 SYNTACTIC ADAPTATION 15 suggest that comprehenders continuously adjust their subjective estimates of the statistics of the input, thus providing support for and extending experience-based accounts of language processing, broadly construed. Syntactic Adaptation Variability in the linguistic environment. Above we summarized evidence for the lack of invariance in the mapping between perceptual data and phonetic categories. Is there a similar lack of invariance at the syntactic level? Intuition suggests that perhaps the answer is no : idiosyncratic variability in the low-level, perceptual qualities of individual talkers speech seems anecdotally obvious (the phonetic level is often the locus of very salient variability along the dimensions of dialect, class, gender, etc.). At the syntactic level, on the other hand, the presence of such variability across and within talkers is much less obvious. Nevertheless, recent work suggests there is considerable variability in how the syntactic structures in a given language are distributed. First, extensive work in sociolinguistics has shown variability across geographical regions, time, modalities and genres in the distribution of syntactic structures. For example, Weiner and Labov (1983) find that the choice between active (Bob killed Laura Palmer) and passive (Laura Palmer was killed by Bob) voice is influenced by the age, gender, and social status of the talker. Similarly, Tagliamonte and Smith (2005) examined the factors governing speakers decision to use the optional complementizer that (e.g., Windom Earle said that he was coming to Twin Peaks), and found that the factors contributing to the mentioning of that, varied across situations (see also Tagliamonte, D Arcy, & Jankowski, 2010). Similarly, Bresnan and Hay (2008) found

16 SYNTACTIC ADAPTATION 16 that syntactic preferences in the dative alternation (such as the choice between (3) and(4)) vary depending on the variety of English used by the speaker (e.g., New Zealand vs. American English). (3) The giant brought Agent Cooper a message. (4) The giant brought a message to Agent Cooper. Providing further evidence for the existence of non-trivial variability at the syntactic level, work in psycholinguistics demonstrates that the way in which syntactic structures are distributed also varies below the level of dialect, region, and other social categories between and even within individual speakers. For instance, in an experimental paradigm eliciting sentences that can occur with or without the complementizer that (Windom Earle said (that) he was coming to Twin Peaks), the same stimuli give rise to considerable variation in rates of that-mentioning across subjects (Ferreira & Dell, 2000). Significantly, it is possible that variation between talkers in the rate of that-mentioning at a given point in the experiment depends on what each talker has produced up to that point in the experiment, since speakers can be primed by their own previous productions. In other words, even within individual talkers, the distribution of syntactic structures may vary as a function of time simply because subjects may prime themselves and become more likely to re-use previously used syntactic structures (as classically observed by Bock, 1986). In short, while the lack of invariance has played a particularly central role in theories of speech perception, a similar lack of invariance is observed in syntax. Variability is thus likely to be a general fact about the distribution of linguistic events in a

17 SYNTACTIC ADAPTATION 17 population of speakers. Thus, a comprehension system that places a premium on efficiently processing linguistic input will be benefit from allowing its estimates of the probabilities of various linguistic events to be sensitive to speakers, genre, etc. Because it is simply impossible that a comprehender will encounter an example of every possible situation before some relatively early point in life, and then use these experiences to guide subsequent comprehension, one way in which comprehenders might cope with variability in the input is to adapt their representation of the statistics of the input to new linguistic situations. This kind of adaptation has been demonstrated in phonetic perception (discussed above). Next we summarize previous work in syntactic comprehension providing initial evidence for adaptation of a similar quality. Previous evidence for syntactic adaptation. For phonetic perception, listeners seem to cope with variability by adapting to the distributional properties of novel linguistic situations (such as novel speakers). Does this ability extend to the syntactic level? We reviewed evidence above suggesting that the same types of variability that may motivate adaptation in speech perception are present at the syntactic level. Insofar as adaptation is expected to the extent that it is merited by properties of the environment (cf. Anderson, 1990), then we should expect syntactic processing to display flexibility that is qualitatively similar to that observed in phonetic perception. For syntactic comprehension, the question is whether humans rapidly adjust their beliefs about the distribution of syntactic structures generated in a particular context or by a particular talker.

18 SYNTACTIC ADAPTATION 18 Some evidence that language users possess the ability to quickly learn distributional regularities over more abstract linguistic units such as syllables, words, and syntactic categories comes from the developmental and statistical learning literature (Braine et al., 1990; Gómez & Gerken, 1999; Saffran, Aslin, & Newport, 1996; Saffran, Newport, & Aslin, 1996; Wonnacott, Newport, & Tanenhaus, 2008). For instance, Saffran and colleagues showed that both children and adults can exploit transitional probabilities between syllables in order to segment words in the input. Similarly, numerous researchers using artificial language learning paradigms have found evidence that knowledge of syntax might be acquired on the basis of statistical regularities defined over syntactic and lexical categories (Braine et al., 1990; Fedzechkina, Jaeger, & Newport, 2011; Wonnacott et al., 2008). These results demonstrate that adults maintain a remarkable capacity for learning statistical regularities in a novel, artificial language. The results thus raise the question of whether adults actively engage in a qualitatively similar sort of statistical learning with their native language. More direct evidence that adaptation extends to syntactic comprehension comes from previous work on priming in syntactic comprehension. Syntactic priming was first observed in production: in her seminal paper, Bock (1986) found that, in a situation where speakers could describe a scene using either version of the dative alternation (e.g., (3) and (4) above), the probability of a speaker producing a given structure significantly increased if they had recently produced that same structure. In comprehension, syntactic priming refers to the facilitated comprehension of a syntactic structure after recent exposure to that structure. Considerably less is known about syntactic priming in comprehension compared to production. A spate of studies

19 SYNTACTIC ADAPTATION 19 suggested initially that while syntactic priming is operative in comprehension (cf. Luka & Barsalou, 2005; Noppeney & Price, 2004), it is of a different character than in production, driven by specific lexical items, and therefore not operative over abstract syntactic units (e.g., Arai, van Gompel, & Scheepers, 2007; Branigan, Pickering, & McLean, 2005; Traxler & Pickering, 2005). Recent work has, however, has found evidence for abstract structural priming in syntactic comprehension, mirroring production in this respect (C. Kim, Carbary, & Tanenhaus, submitted; Thothathiri & Snedeker, 2008; Traxler, 2008). However, there are major gaps in our understanding of the relevant phenomena that this paper attempts to address. First, previous work on priming in comprehension has focused exclusively on the effect on comprehension behavior during a target immediately following a prime (Arai et al., 2007; C. Kim et al., submitted; Thothathiri & Snedeker, 2008; Traxler, 2008; Traxler & Pickering, 2005). By contrast, we ask whether priming can be better characterized as a form of learning whereby each syntactic structure that a comprehender encounters counts as a piece of evidence that the comprehender uses to update their beliefs about the distribution over syntactic structures. Second, by construing syntactic adaptation as the incremental update of a distribution over syntactic structures, we derive specific predictions about the shape of the learning function underlying syntactic adaptation. In the Discussion, we consider in more detail the relationship between our proposals and previous work on syntactic priming. Goals and Predictions of the Current Study We propose that comprehenders deal with syntactic variability by continuously adapting to the statistics of the linguistic environment. To test this proposal, we provide

20 SYNTACTIC ADAPTATION 20 subjects with experience with written language in which multiple probabilistic cues to syntactic structure are present, but in which the distributional properties of the environment differ sharply from subjects previous experience with language. If subjects are adapting to the statistics of the input, as we propose, then the manner in which these probabilistic cues influence behavior should change in a principled way over the course of the experiment. In order to concretely address this question, we exploit syntactic ambiguities previously studied in the syntactic processing literature. In Experiment 1, we focus on sentences like (5), which share the sentence complement (SC) vs. direct object (DO) ambiguity with example (1) above. It is disambiguated toward the sentence complement (SC) reading at should, which we refer to as the point of disambiguation. There are at least two cues to syntactic structure in sentences such as these, and these cues are strong predictors of reading times at the point of disambiguation. The first is whether the complementizer that is present, providing an early disambiguating cue (e.g., The CIA director confirmed that the rumor should have been stopped sooner); when present, reading times at the point of disambiguation are relatively low, since comprehenders have a strong expectation for an SC. The second cue is the verb itself some verbs are more likely than others to occur with SCs; we will refer to this property of verbs as SC-bias (formally, p(sc v), or the probability of an SC given a verb, v). Researchers can estimate SC-bias either from language corpora or from norming studies where participants complete sentence fragments such as The CIA director confirmed. The proportion of SC completions provide an estimate of the average English speaker s implicit subjective beliefs about how likely an SC vs. a DO is to occur with this subject noun phrase-verb pair.

21 SYNTACTIC ADAPTATION 21 (5) The CIA director confirmed (that) the rumor should have been stopped sooner. Numerous studies demonstrate that comprehenders use SC-bias to constrain their interpretation of the unfolding sentence, and the effect of SC-bias is visible at the point of disambiguation the more likely a verb is to take an SC, the lower reading times will be at the point of disambiguation, all things being equal (Garnsey et al., 1997; Trueswell et al., 1993). Crucially, the presence or absence of the complementizer that and SC-bias interact when the complementizer that is present, thereby providing a strong, early disambiguating cue, the verb bias has less of an effect on behavior. One way of thinking about this tradeoff appeals to the notion of cue reliability, classically defined by Bates and MacWhinney (1987) as how often the use of a cue leads to the correct inference. When multiple cues are available to guide an inference, rational models of perception predict that humans will depend more on cues that are more reliable; more precisely, these models predict that cues should be weighted according to the inverse of their variance, where the variance of a cue is one way to quantify its reliability (Alais & Burr, 2004; Ernst & Banks, 2002; Jacobs, 2002). In the case of temporary ambiguities like the one at the rumor in (5) inferences or predictions about syntactic structure will be guided by the verb and the complementizer that, when present. If that is present, subjects depend on this cue a great deal more than the verb itself because that is a much more reliable cue to syntactic structure when the word that follows a SC-taking verb, the

22 SYNTACTIC ADAPTATION 22 probability of an SC structure is essentially 1. 1 This tradeoff surfaces in a regression framework as a two-way interaction between the complementizer that and the bias of the verb, previously observed by numerous researchers (e.g., Garnsey et al., 1997; Trueswell et al., 1993). However, in our experiments, the statistics of the linguistic environment differ sharply from subjects a priori expectations (as captured by, say, norming statistics). Specifically, in our first experiment, all items are disambiguated towards the SC structure. Thus, a comprehender adapted to the statistics of this specific environment should, over the course of the experiment, come to assign a much higher probability to SCs than they otherwise should. This leads to a concrete prediction about the relationship between reading times, a priori verb statistics, and the point in the experiment at which an item is encountered. Above we explained that the effect of SC-bias on reading times surfaces as a two-way interaction between SC-bias and the presence vs. absence of that, but if subjects are adapting to the statistics of the environment (i.e., of the experiment), then the 1 For present purposes, it is fair to say that p(sc that ) = 1. However, there are at least two reasons why this is, strictly speaking, inaccurate. First, the word that is lexically ambiguous between the complementizer interpretation (as intended in our materials) and a determiner or pronoun interpretation (e.g., Audrey believed that agent with the bullet-proof vest or Audrey said she saw the agent but nobody believed that). Second, perception is noisy, so that comprehenders do not have certainty as to what they have perceived. A rational comprehender should preserve this uncertainty and indeed there is evidence that comprehenders maintain uncertainty about previous input (so-called right context effects in speech perception, see 1/12/12 9:38 AM) for an overview; for similar effects in syntactic processing, see 1/12/12 9:38 AM; for rational observer models of these effects, see 1/12/12 9:38 AM.

23 SYNTACTIC ADAPTATION 23 relationship between reading times and the a priori (i.e., pre-experimental) SC-bias of the verbs should be dampened in that linguistic environment. Therefore, we predict a threeway interaction between SC-bias, the presence vs. absence of the complementizer that, and when in the experiment an item is encountered by the subject, a variable we will refer to as item order. The remainder of this paper is organized as follows: In the following section we present the results from a self-paced reading experiment providing evidence that humans rapidly adapt to the statistics of the input. We then present a second experiment that extends this result and addresses some questions that arise with Experiment 1. Next we discuss these, paying special attention to the similarities and differences between our findings and evidence for adaptation at the phonetic level and in non-linguistic domains, as well as to the relationship between our findings and previous work on syntactic priming. Experiment 1 In our first experiment, we conduct a modified version of a classic sentence processing experiment (Garnsey et al., 1997) in which subjects read temporarily ambiguous sentences such as (5). Following Garnsey et al. (1997), all sentences in the experiment containing DO/SC verbs occurred with SCs; therefore, the local, experimentspecific statistics of the input drastically diverge from the statistics of everyday language, where SCs occur far less frequently. As explained above, if subjects rapidly adapt to the statistics of the experiment, then their subjective estimate of the probability of an SC should increase sharply over the course of the experiment, and we expect this to surface as a three-way interaction between (1) whether the sentence is ambiguous (presence vs. absence of the complementizer that); (2) the SC-bias of the verb; and (3) item order.

24 SYNTACTIC ADAPTATION 24 Subjects 94 subjects were recruited from the University of Rochester undergraduate community. All subjects had normal or corrected to normal vision. Informed consent was obtained from all subjects according to the guidelines of University of Rochester Research Subjects Review Board. Subjects were paid $10.00 compensation for their participation. The data from one subject was excluded when it was discovered after the experiment that the subject was not a native speaker of English. Materials Sentence stimuli consisted of 36 items and 72 fillers. The 36 items comprise a subset of those used in (Garnsey et al., 1997). Items were manipulated along two dimensions: between items, verbs were chosen such that the conditional probability of an SC, given the verb, varied. In Garnsey et al. (1997), this conditional probability was estimated from a norming study, in which subjects completed sentence fragments containing DO/SC verbs (e.g., The reviewers acknowledged...). The verbs used in the critical sentences were selected to span a wide range of SC-bias values, from.01 to.9, as reported in Garnsey et al. (1997). Within items, we manipulated complementizer presence: half of the sentences presented to each subject included the complementizer that. Garnsey et al. (1997) also manipulated the plausibility of the post-verbal NP (the rumor in (5) above, repeated below), to either bias towards a DO or an SC interpretation, as determined by the norming study. Here, we used only the subset of the items with post-verbal NPs biasing towards the DO interpretation. That is, our items are expected to elicit large garden path effects.

25 SYNTACTIC ADAPTATION 25 (6) The CIA director confirmed (that) the rumor should have been stopped sooner. Items were constructed such that the subject noun phrase (e.g., The CIA director) always contained three words, the post-verbal noun phrase (e.g., the rumor) always contained two words, and the remainder of the sentence after the post-verbal NP contained five words (e.g., should have been stopped sooner). Filler sentences contained a variety of syntactic structures, including 32 intransitives (e.g., The shoe salesman looked impatiently out the window while waiting for customers), 10 passive constructions (e.g., The university library was unlocked by the head librarian early in the morning), and 30 simple transitives (e.g., The varsity cheerleader cried uncontrollably when the home team lost). Two lists were constructed such that each item occurred once in each list and the presence/absence of the complementizer that is counterbalanced across lists. Subjects were randomly assigned to one of these lists, and randomly assigned a presentation order. Procedure Subjects read sentences in a self-paced moving window display (Just et al., 1982) presented using the Linger experimental presentation software (Rohde, 2005). At the beginning of each trial, the sentence appeared on the screen with all characters replaced by a dash. Subjects pressed the space bar using their dominant hand to view each consecutive word in the sentence. Durations between space bar presses were recorded. At each press of the space bar, the currently-viewed word reverted to dashes as the next word was

26 SYNTACTIC ADAPTATION 26 converted to letters. A yes/no comprehension question followed all experimental and filler sentences, with the correct answer to half of all comprehension questions being yes. Analysis Because one of the strongest predictors of reading times (RTs) in reading studies is word length (i.e., number of characters per word), it is standard to remove the variance accounted for by word length in RTs to yield the conventionally used length-corrected RTs. Before computing length-corrected reading times, all raw RTs that were abnormally low (i.e., below 100 ms) or abnormally high (i.e., above 2000 ms) were removed, resulting in 0.1% data loss. Trials on which subjects incorrectly answered the comprehension question were removed, resulting in 7% data loss. Length-corrected RTs were computed by regressing the remaining raw RTs onto word length using linear mixed effects regression. In addition to a single main effect of word length, this model included a random intercept for subject, and a by-subject random slope for length (these random effects allow the model to discount mean differences in raw RT across subjects as well as variable sensitivity to the effect of word length across subjects). The residuals of this model, length-corrected RTs, served as the dependent variable in all analyses reported below. Though subjects read all sentences word-by-word, for the purposes of analysis items were divided into five separate regions, shown in (7). As is standard for this type of analysis, the disambiguation region consisted of the first word that unambiguously enforced the SC reading and the following word. This is done because reading difficulty at the point of disambiguation often spills over onto the next word(s) (Mitchell, 1984).

27 SYNTACTIC ADAPTATION 27 sooner. (7) The CIA director / confirmed / that / the rumor / should have / been stopped We focus on per-word length-corrected RTs at the disambiguating region. Note that when the complementizer that is present, the sentence is, in fact, disambiguated before what we are calling the disambiguating region. Nevertheless, we use this label across sentences both with and without the complementizer for ease of exposition. Length-corrected RTs during the disambiguating region (should have in (7)) were regressed onto the full factorial design (all main effects and all interactions) of the following predictors: (a) log-transformed verb bias, (b) complementizer presence (that vs. no that) and (c) item order. Item order was coded as an index, 1 36, of the point in the experiment at which the experimental item was presented relative to other experimental items. Recall that item order corresponds to the point in the experiment at which an item is processed and therefore captures how much evidence the subject has accrued concerning the distribution of SCs, up to that point in the experiment. As an additional control, we included (d) a main effect of log-transformed stimulus order. Stimulus order differs from item order in that it is an index of when the item was presented relative to both items and fillers. Stimulus order therefore captures essentially how long subjects have been doing the experiment, whereas item order captures how many SCs subjects have observed. The reason for including log stimulus order was to control for task adaptation an overall speed-up in reading times (across both fillers and items and in all sentence regions) over the course of the experiment arising purely from

28 SYNTACTIC ADAPTATION 28 increasing familiarity with the task itself; previous work suggests that task adaptation is best captured by a log-linear relationship between RTs and order. Additionally, the model included the maximum random effects structure justified by the data based on model comparison using log-likelihood ratio tests (for discussion, see Barr, Levy, Scheepers, & Tily, submitted; Jaeger, 2009; Jaeger, Graff, Croft, & Pontillo, 2011). A detailed description of the procedure used to determine the fixed and random effects structure for the model is included in Appendix A1 and A.2. All predictors were centered at zero to reduce collinearity with interactions. Results Experiment 1 replicated previous studies exploiting the same ambiguity, such as Garnsey et al. (1997) on whose materials those of the current study were based as well as Trueswell et al. (1993). First, there was a significant main effect of complementizer presence: length-corrected RTs at the disambiguating region were lower when that was present (β = 7.9; p<.05). The main effect of verb bias where length-corrected RTs decrease as the SC-bias of the verb increases was marginally significant (β = 6.3; p =.1). Importantly, though, these two variables interacted as expected verb bias had a diminished effect when the complementizer that was present (β = 5.7; p <.05). This interaction is visualized in the top two panels of Figure 2. The bottom panels, taken from Garnsey et al. (1997), show the same interaction from the Garnsey et al. study. On the y- axis we show the difference score between reading times, at the point of disambiguation, between sentences without the complementizer that and those with. As mentioned above,

29 SYNTACTIC ADAPTATION 29 this is referred to as the ambiguity effect. The ambiguity effect is plotted against SC-bias on the x-axis. No other two-way interactions reached significance XXXXXXX Insert Figure 2 here XXXXXX The central prediction for the current study is that, as subjects observe more and more SC structures, their subjective beliefs about the conditional probability p(sc verb, that) should adapt to the statistics of the experiment. That is, we expected a three-way interaction between verb bias, the presence of the complementizer that, and item order, which captures the number of observations of the SC structure at a given point in the experiment. This three-way interaction was indeed significant in the expected direction (β = -.48; p <.05) as the experiment progressed, the processing advantage conferred by verbs with a high a priori probability of occurring with an SC decreased. To visualize changes in subjects subjective beliefs about p(sc verb, that), we therefore plot in, Figure 3, how the correlation coefficient between the ambiguity effect and verb bias changes across the experiment. As is apparent in Figure 3, the initially negative relationship becomes more and more positive over the course of the experiment. This change in the direction and magnitude of the ambiguity effect is predicted by the hypothesis that sentence comprehension is an adaptive process in the sense assumed here XXXXXXX Insert Figure 3 here XXXXXX