DOMINANCE AND CONTEXT EFFECTS 1 Running Head: DOMINANCE AND CONTEXT EFFECTS ON HOMOPHONE MEANINGS Dominance and context effects on activation of alternative homophone meanings

Two eye-tracking during listening experiments found frequency and context effects on fixation probability of pictures representing multiple meanings of homophones. Participants heard either an imperative sentence instructing them to look at a homophone referent (Experiment 1) or a declarative sentence that was either neutral or biased towards the homophone‘s subordinate meaning (Experiment 2). At homophone onset in both experiments, participants viewed four pictures: 1) a referent of one homophone meaning, 2) a shape competitor for a non-pictured homophone meaning, and 3) two unrelated filler objects. In Experiment 1, meaning dominance affected looks to both the homophone referent and the shape competitor. In Experiment 2, compared to neutral contexts, subordinate-biased contexts lowered the fixation probability for shape competitors of dominant meanings, but shape competitors still attracted more looks than chance. We discuss the consistencies and discrepancies of these findings with the Selective Access and Reordered Access theories of lexical ambiguity resolution. DOMINANCE AND CONTEXT EFFECTS 3 Intuition suggests that we normally settle on the appropriate meaning of a lexically ambiguous word immediately if the context is sufficiently informative, e.g., Chris needed some money, so he went to the bank. In fact, a large number of experimental studies have demonstrated that sentence context influences the timing and/or degree to which the alternative meanings are activated (e.g., Dopkins, Morris, & Rayner, 1992; Lucas, 1999; Sereno, Brewer & O‘Donnell, 2003, Tabossi, 1988). However, the relative frequency of the alternative meanings is also clearly important. When presented in a neutral sentence context, the more frequent, dominant meaning of an ambiguous word is activated more quickly and persists longer than less frequent, subordinate meanings (e.g., Simpson & Burgess, 1985; Simpson & Krueger, 1991). Most of the current debate in the domain of lexical ambiguity resolution rests on the fate of the dominant meaning when the context supports the subordinate meaning. The Reordered Access account of Duffy, Morris, and Rayner (1988) maintains that both the subordinate meaning (activation boosted by context) and the dominant meaning (activated due to its frequency) compete for selection. Thus, in contexts supporting subordinate meanings, reading times for ambiguous words are slow, compared to an unambiguous control word that is matched in corpus frequency to the overall frequency of the ambiguous word. This is the ―Subordinate Bias Effect‖ that has been reported in numerous reading experiments (Binder, 2003; Binder & Rayner, 1998; Kambe, Rayner & Duffy, 2001; Pacht & Rayner, 1993; Rayner, Cook, Juhasz, & Frazier, 2006; Rayner, Pacht & Duffy, 1994, Sereno, O'Donnell, & Rayner, 2006; Sereno, Pacht & Rayner, 1992). The Subordinate Bias Effect is a well-established phenomenon under two conditions: (1) the homograph is strongly polarized, with subordinate meanings retrieved only about 10% of the time in word association tasks; (2) reading time for the homograph in subordinate context is DOMINANCE AND CONTEXT EFFECTS 4 compared to an unambiguous word matched to the homograph‘s form frequency—i.e., the sum of all meaning frequencies. When these conditions are not met, the results have been less consistent. For example, Wiley and Rayner (2000) did not find a Subordinate Bias Effect, using ambiguous words that were not strongly polarized and context passages that made use of titles for disambiguation. And Sereno et al. (2006, Experiments 2 and 3) actually found the reverse pattern when using a meaning-frequency control word: reading times were faster for highly polarized homographs in strong subordinate-biased contexts, compared to a control word that was matched in frequency with the subordinate meaning of the homograph. This finding seems to suggest that only the subordinate meaning of the homograph was accessed, but it is unclear why the homographs were faster than the control words—the authors themselves simply speculate and recommend further research. Sereno et al. (1992), who also failed to find the standard Subordinate Bias Effect, argued that a meaning-frequency control word is more appropriate than a form-matched control word, because the latter contrasts a high frequency (control) word with a low frequency word (the subordinate meaning of the homograph). The issue of the appropriate control word is critical, although it is difficult to resolve because both word form frequency and word meaning frequency are likely to impact reading time. Furthermore, the fact that the Subordinate Bias Effect is so dependent on comparison against a high frequency control word raises the concern that the Subordinate Bias Effect does not reflect competition between meanings after all, but rather the increased time it takes to access and integrate a lower frequency word (as in Reichle, Pollatsek, & Rayner, 2006; Reichle, Rayner, & Pollatsek, 2003). Reichle et al. (2006) suggest that the Subordinate Bias Effect can be modeled in two ways. In their preferred account, slower reading times on the subordinate-biased homograph are due to competition from the dominant meaning. Because both meanings are DOMINANCE AND CONTEXT EFFECTS 5 activated and in competition, this is consistent with the Reordered Access model (Duffy, Kambe & Rayner, 2001; Duffy et al., 1988). Reichle et al. (2006) also consider an account in which slower reading times for the subordinate-biased homograph are due to its low frequency, compared to the high frequency control word. Such an account is consistent with selective activation of the subordinate meaning only, because it provides an explanation of the Subordinate Bias Effect without reference to activation of the dominant meaning. During the last 15 years, Kellas and colleagues have advocated for a Context-Sensitive or Selective Access account in which the dominant meaning is not activated if the context is sufficiently constraining toward a subordinate meaning. Evidence from lexical decision, naming and self-paced reading studies has demonstrated that in strongly biasing contexts, reaction times to contextually appropriate items are facilitated, but reaction times to contextually inappropriate meanings are not (Kellas, Martin, Yehling, Herman, & Vu, 1995; Martin, Vu, Kellas & Metcalf, 1999; Kellas & Vu, 1999; Simpson, 1981; Vu & Kellas, 1999; Vu, Kellas, Metcalf & Herman, 2000; Vu, Kellas & Paul, 1998; Vu, Kellas, Petersen & Metcalf, 2003). In short, the fate of the dominant meaning in a strong subordinate context is a theoretically important question that differentiates two possible accounts of the Subordinate Bias Effect, and more generally, distinguishes between the Reordered and Selective Access accounts of lexical ambiguity resolution. The issue is whether top-down contextual cues can override the strong relationship between the word-form of an ambiguous word and its dominant meaning. This issue applies to both reading and listening paradigms. In fact, much of the ground-breaking research used spoken homophones in a cross-modal paradigm, leading to the conclusion that multiple meanings are accessed, even in biasing context (e.g., Onifer & Swinney 1981; Swinney, 1979; Tabossi, Colombo & Job, 1987; Tabossi, 1993; Tanenhaus, Leiman, & Seidenberg, 1979). DOMINANCE AND CONTEXT EFFECTS 6 More recently, Huettig and Altmann (2007) found evidence that the dominant homophone meaning is activated in a subordinate-biased context, using a variant of the visual world paradigm introduced by Tanenhaus, Spivey-Knowlton, Eberhard, and Sedivy (1995). We take some time to review Huettig and Altmann‘s (2007) second experiment, both because of its similarity to the current set of studies and because it provides quite dramatic evidence of dominant meaning activation in subordinate context. Participants heard sentences containing polarized homophones, such as ―pen‖ while viewing an array of four line drawings. The drawings appeared one second before sentence onset in all conditions. In the neutral condition, the sentence contexts were neutral to the meaning of the homophone, such as The man got ready quickly, but then he checked the pen. In the biased condition, the sentence contexts supported the subordinate meaning of the homophone, e.g., The welder locked up carefully, but then he checked the pen. In the neutral and biased conditions (25% of the trials), participants saw drawings depicting both meanings, along with two unrelated distracters. In the competitor condition, participants heard the subordinate-biased context sentence while a visual shape competitor (a sewing needle) replaced the dominant meaning (pen-writing) (for evidence that looks to visual shape competitors track lexical access, see Dahan and Tanenhaus, 2005; Huettig, Gaskell, & Quinlan, 2004). At homophone onset (before lexical access of the homophone had begun), Huettig and Altmann (2007) found no differences in looks to any objects in the neutral condition. However, in the biased and competitor conditions, the listeners were already fixating the pen-enclosure drawing 45% and 49% of the time, respectively, indicating that they had used the sentence context to fixate on sentence-relevant drawings. DOMINANCE AND CONTEXT EFFECTS 7 At homophone offset, both dominant and subordinate referents were fixated more than fillers in both the neutral and the biasing conditions. In the neutral condition, presumably both meanings of the homophone were accessed due to support from two sources: the spoken word form and the images of two potential referents in the visual display. It is not possible to know for sure that meaning activation for the homophone was influenced by the visual display in this condition, but it seems likely, especially for the subordinate meaning. In the biased condition, the combination of sentence context and the visual display apparently activated the subordinate meaning prior to any phonological cues from the spoken homophone. Nonetheless, it is striking that the spoken word-form of the homophone still prompted looks to the dominant, but contextually inappropriate, referent. Even more striking is the fact that the pattern for the competitor condition was very similar to that in the biased condition, even though the dominant referent was replaced with a shape competitor. Looks to the shape competitor for the dominant meaning can be taken as evidence that the dominant meaning was activated, despite the subordinate context and the lack of an appropriate visual referent. Thus, this experiment provides very strong evidence for dominant meaning activation, even when a subordinate-biased context has successfully focused attention on the subordinate meaning. The Huettig and Altmann (2007) findings provide compelling evidence that the wordform of a polarized homophone will always activate the dominant meaning, regardless of the linguistic and visual context. This begs the question: does the linguistic context have any effect on the activation of the dominant meaning? According to Reordered Access, a subordinatebiased context can boost activation of the subordinate meaning, thereby making the dominant meaning and subordinate meaning more competitive. But this does not entail delaying and/or limiting activation of the dominant meaning. The Reordered Access model maintains that ―prior DOMINANCE AND CONTEXT EFFECTS 8 disambiguating context does affect the access process by increasing the availability of the appropriate meaning without influencing the alternative meaning‖ (Duffy et al., 1988, p. 431). The same point has been made more recently: ―The two meanings became activated independently. While context could speed activation of the intended meaning, it had no effect on the speed of activation of the unintended meaning‖ (Rayner, Binder & Duffy, 1999, p. 847). Thus, in subordinate-biasing contexts, activation of the dominant meaning should be unaffected, while the subordinate meaning should be activated earlier than usual. On the other hand, for Selective Access accounts, the activation of the meanings of an ambiguous word depends on several constraints: frequency, type of context, and strength of context; the combined influence of these variables determines the meaning accessed (Martin et al., 1999). Thus, the subordinatebiasing context would serve to both boost activation of the subordinate meaning and limit activation of the dominant meaning. In the current experiments, we use a visual world paradigm similar to that of Huettig and Altmann (2007), and begin by establishing the extent of meaning activation for subordinate and dominant meanings of homophones in neutral context (Experiment 1). Then, we manipulate the linguistic context to determine how activations of both the subordinate and dominant meanings are affected (Experiment 2). As in Huettig and Altmann‘s competitor condition, we used visual displays that contained an actual referent for the subordinate meaning, a shape competitor for the dominant meaning, and two unrelated distractors. One major difference between our experiments and theirs is that we do not present the visual display until homophone onset, in order to prevent the visual display from influencing the initial stages of spoken word recognition. Experiment 1 DOMINANCE AND CONTEXT EFFECTS 9 Experiment 1 explores dominance effects on the activation of multiple meanings of ambiguous words in an instructional eye-tracking during listening task. This is somewhat analogous to the neutral condition of Huettig and Altmann (2007), described above, but there are three important differences. First, instead of using a declarative sentence, we used interrogatives that directly engaged the listener (―Look at the flower/flour.‖) in the tradition of Dahan, Magnuson and Tanenhaus (2001). Second, we presented the visual stimuli coincident with the onset of the target word (a homophone in our experiments), rather than at trial onset, in order to limit the degree to which the visual context constrains lexical activation. Our goal in making these changes was to get a clearer picture of the time-course and extent of activation for the dominant and subordinate meanings, as reflected by fixation patterns in this paradigm, when meaning frequency is the only relevant factor. Third, rather than presenting two meanings of the polarized homophones directly, one meaning was depicted using an actual referent picture and the other meaning was depicted by a visual shape competitor (Dahan & Tanenhaus, 2005, Huettig et al., 2004). For example, the dominant meaning of the homophone flower/flour was directly depicted using a picture of a flower, and the subordinate meaning was indirectly depicted using a pillow as a visual shape competitor for flour. On another trial, the subordinate meaning of flower/flour was depicted using a picture of flour, and a lollipop was used as a shape competitor for the dominant meaning, flower. Although an actual referent to the homophone was always in the visual display, residual activation of the alternative meaning of the homophone could be assessed by analyzing the looks received by the shape competitor. As in Dahan and Tanenhaus (2005), none of the shape competitors overlapped in phonology with the spoken referent names, so any activation of the shape competitor from the spoken input indicated activation of its homophone referent. The DOMINANCE AND CONTEXT EFFECTS 10 activation of multiple meanings of the ambiguous word was measured by comparing looks to the shape competitor picture on trials where the dominant or the subordinate actual item was pictured. A relative dominance effect would be found if pictures of dominant meanings of a homophone attracted more looks than pictures of subordinate meanings. Because we needed to compare the probability of looking at two different pictures, it was crucial to match the pictures on various dimensions, as described in our normative measures below. Method Participants Thirty undergraduates at the University of Michigan participated in this experiment for partial course credit in an Introductory Psychology class. All participants in this and the following experiments were native speakers of English and had normal or corrected vision. Materials We collected word association norms for a set of heterographic and homographic homophones (details below) and chose the 14 homophones for which meaning dominance was most polarized. Each homophone had two distinct, imageable meanings (see Appendix A). The auditory stimuli were recorded by a female speaker: ―Look at the cross. Now look at the (homophone).‖ For each digital speech file, silence was added before the onset of the spoken instructions as needed (i.e., before ―Look at...‖), so that the onset of each critical homophone was 3000 ms from the beginning of the auditory stimulus. For each meaning of each homophone, two critical pictures were selected. One depicted the referent of the homophone (Actual Referent), and one depicted an object that was similar in shape to the homophone referent (Shape Competitor). The norms used to assess shape similarity are described below. DOMINANCE AND CONTEXT EFFECTS 11 Visual stimuli consisted of four pictures arranged on a white background with a fixation cross in the center for each critical trial. The critical pictures were all full-color realistic images selected from an online searchable database of images (Google, 2004). Actual Referent images were chosen so that the picture represented a typical instance or instances of the object, and such that the picture would not be identified with other possible labels. For example, a stamp with an unknown design was chosen so that the participants would not identify the stamp with its design, such as an ―American flag‖. In the case of letter, the Actual Referent included multiple letters in order to prevent the picture to be labeled as the letter itself, such as ―B‖. Shape Competitor images were selected so that the picture was as identifiable as the Actual Referent and would not be assigned a label that overlapped in phonology or semantics. The pictures appeared in the upper left, upper right, lower left, or lower right quadrant, arranged so that the Actual Referent appeared in each quadrant an equal number of times for every participant. The Shape Competitor also appeared in each quadrant an equal number of times for every participant. The two remaining quadrants contained filler pictures of objects with unambiguous names that did not begin with the same phonemes as the critical homophone and were not similar in shape to the critical pictures. Each homophone was tested only once, with each auditory stimulus occurring with one of two visual display types: Dominant-Actual or Subordinate-Actual. In the Dominant-Actual display, participants viewed the Actual Referents of the dominant meaning of the homophones, together with the Shape Competitors of the subordinate meaning of the homophones. In the Subordinate-Actual display, participants viewed the Actual Referents of the subordinate meaning of the homophones, together with the Shape Competitors of the dominant meaning of the homophones. Display type was varied between participants because there was awe were DOMINANCE AND CONTEXT EFFECTS 12 concerned that possibility that Shape Competitor effects would not be large enough to create a comparisons would get relatively few looks, and we wanted to maximize statistical power for comparisons evaluating the hypothesis that Shape Competitors received more looks than would be expected by chance. Each participant was randomly assigned to one of the two display types. In addition to the 14 critical homophone trials, 14 filler trials with unambiguous targets were constructed. Trials were presented in a fixed random order. Norming Word association norms. Meaning dominance frequencies were collected in a word association task. Twenty-seven participants provided the association norms and received partial course credit. No participants in this and any of the norming experiments participated in the primary experiment. Participants listened to a list of words and, for each word, typed the first related word that came to mind. The stimuli consisted of 148 heterographic and homographic homophones and 80 unambiguous fillers. We selected 14 homophones that elicited word association responses with at least a 19% difference between the dominant and subordinate meaning. Of these 14 homophones, on average, the dominant meaning gathered 79% of the total word association responses, and the subordinate meaning gathered 16% of the total responses. The remaining 5% of the responses had missing values or were unrelated to the two most common meanings of the homophone. Picture agreement norms on Actual Referents. Picture agreement norms confirmed that there was no difference in labeling agreement for the pictures we chose to represent the Actual Referents of the subordinate and dominant meanings of the homophones. Forty-two participants were each presented with a sequence of 200 pictures on a computer screen. Two lists were created such that half of the homophone pictures were of the dominant meaning, and half of the DOMINANCE AND CONTEXT EFFECTS 13 homophone pictures were of the subordinate meaning. Each participant saw each homophone item once, either in the dominant or subordinate condition. Only one picture appeared on each screen, simultaneously with a box in which they were asked to type the name of the picture of the object represented. Fourteen were pictures of homophones, and 186 were filler pictures with unambiguous labels. Trials were coded as having correct agreement when the response included the homophone in any part of the answer, including misspelled words and plurality differences, but not including synonyms or other names. For example, if the intended label was flower, responses such as purple flower, flower petals, flowers, and flowr were accepted. However, responses such as orchid, purple bloom, and bouquet were not accepted. The agreement between the participants‘ responses and the intended labels was 85.5% for the dominant-biased condition and 81.2% for the subordinate-biased condition. A t-test indicated no differences between conditions (t2(26) = 0.76, p > .10). Picture norms on Shape Competitors. In choosing the pictures to represent the Shape Competitors, it was not crucial to select pictures with high name agreement, because the relevant factor was shape similarity to a prototypical object representing one of the homophone‘s meanings. Nonetheless, it would present a confound if a shape competitor picture was likely to be labeled with a name beginning with the same phonemes as the homophone in that particular trial. This is actually a potential weakness for Huettig and Altmann‘s (2007) competitor condition, described above, because they provided no norms on the shape competitors to ensure that they would not be given a label that overlapped in phonology with the homophone. We worried that, for example, the shape competitor ̳needle‘ from their example sentence, would activate ―pin‖—a phonological competitor for ―pen‖. DOMINANCE AND CONTEXT EFFECTS 14 Thus, we collected picture naming data for the shape competitor pictures, similarly to the labeling agreement norms for the Actual Referent pictures. Twenty participants were presented with a sequence of 128 pictures on a computer screen. One list was created with 14 competitors of the dominant meanings and 14 competitors of the subordinate meanings of the homophone. One hundred filler pictures with unambiguous names were used. Each participant saw every Shape Competitor once, both in the dominant and subordinate conditions. Only one picture appeared on each screen, simultaneously with a box in which they were asked to type the name of the picture of the object represented. No responses for the Shape Competitor pictures indicated any phonologically similarity to the homophone to which the Shape Competitor belonged. Picture similarity norms. Picture similarity norms indicated that the Shape Competitors were in fact similar in shape to the Actual Referents. Twenty-four participants were presented with a series of pictures with questions, such as ―How similar in shape is this object to a flower?‖ Participants were asked to rate on a scale of 1 (not similar) to 7 (very similar) how similar the presented picture was to the object mentioned in the question. Participants‘ judgments may also have been influenced by other perceptual variables, such as color and texture, but we explicitly avoided pictures with any conceptual or instrumental relationships to the homophones. In addition to the 28 Shape Competitor trials, there were 25 filler trials in which participants were asked to evaluate the shape similarity of a picture to a concept that we judged to be either related or unrelated in shape. Participants rated every Shape Competitor once. The mean ratings for shape-similarity were 4.84 and 5.76 for dominant and subordinate Shape Competitor, respectively. T-tests revealed that dominant Shape Competitors were judged less similar in shape to the actual objects than subordinate items, so any looks to dominant Shape Competitors can not DOMINANCE AND CONTEXT EFFECTS 15 be due to higher similarity of those items to the actual objects (t1(1,23) = 7.32, p < .001; t2(1,26) = 3.50, p < .05). Picture saliency norms. Picture saliency norms were conducted to evaluate any difference in saliency among the critical pictures. These norms were collected for all 16 critical trials used in Experiment 2, however only the 14 trials relevant to Experiment 1 are reported here. Thirtyfive participants were asked to view pictures silently on the computer screen while a headmounted eye-tracker monitored their eye movements. Two display conditions were created such that half the participants saw the dominant Actual Referents and subordinate Shape competitor, and the other half saw the subordinate Actual Referents and the dominant Shape competitor. On each trial, four pictures appeared on the screen for 5 seconds, and a drift correction procedure was conducted between every trial. The four pictures were the same as the pictures from Experiment 1. The critical trials and 28 filler trials were presented in a random order. The dwell time percentages for each object type in the Dominant-Actual display condition were the following: Actual Referent: 19.1%, Shape Competitor: 20.1%, Fillers: 18.7%. Subordinate-Actual display: Actual Referent: 21.2%, Shape Competitor: 18.9%, Fillers: 18.5%. For Shape competitors and Fillers, independent t-tests found no differences between the Dominant Actual and Subordinate Actual display conditions (Shape Competitor: t1(1,33) = 1.04, p >.10, t2 < 1; Filler: t1<1, t2 < 1). For Actual Referents, there was a marginal effect only by participants (t1(1,33) = -1.83, p <.10, t2 < 1), in which subordinate referents attracted more fixations. Independent t-tests found no advantage of Shape Competitors compared to Fillers in the Dominant Actual display (t1(1,52) = 1.17, p >.10, t2 < 1) nor in the Subordinate Actual display (t1 < 1, t2 < 1). Thus, there were no advantages in saliency for pictures representing the DOMINANCE AND CONTEXT EFFECTS 16 dominant meanings over pictures representing the subordinate meanings, nor for Shape Competitors over Filler objects. Procedure and Equipment The auditory sentences and their corresponding slides were presented in a fixed random order. There were four practice trials before the experimental trials began. Eye position was recorded as participants listened to the sentences, using an ISCAN© Head-mounted Eye Tracking System. The eye and scene cameras were mounted on a headgear, with a sampling rate of 120 Hz. Participants were seated at approximately 24 inches from the screen. The visual angle from the fixation cross to the pictures was approximately 9 degrees. The auditory and visual stimuli were presented using E-Prime software. Participants heard these instructions: At the beginning of each trial, you will see a cross in the center of the screen. Surrounding the cross, there will be four pictures. You will hear instructions that will ask you to look at the cross and then point to objects on the screen. Before the practice trials, a six-point calibration slide was presented. On each trial, participants were presented with a fixation slide simultaneously with auditory sentence instructions. At 3000 ms after sentence onset, the four-picture slide appeared simultaneously with the onset of the critical homophone. The experimenters used the scene camera screen to verify whether or not the participant was accurately pointing to the correct targets. Between each trial, there was an additional six-point calibration slide. If four out of six points were not accurately calibrated, recalibration was performed. The entire experiment lasted less than thirty minutes. The data were collected and organized using PRZ analysis software provided by ISCAN©. DOMINANCE AND CONTEXT EFFECTS 17