Priming ditransitive structures in comprehension
Manabu Arai University of Edinburgh
Roger P.G. van Gompel University of Dundee
Christoph Scheepers University of Glasgow
Address correspondence to:
Manabu Arai Psychology School of Philosophy Psychology and Language Sciences University of Edinburgh 7 George Square Edinburgh EH8 9JZ Tel: +44 131 650 3410 Fax: +44 131 650 3461 Email:
[email protected] 1
Abstract
Many studies have shown evidence for syntactic priming during language production (e.g., Bock, 1986). It is often assumed that comprehension and production share similar mechanisms and that priming also occurs during comprehension (e.g., Pickering & Garrod, 2004). Research investigating priming during comprehension (e.g., Branigan, Pickering, & McLean, 2005; Scheepers & Crocker, 2004) has mainly focused on syntactic ambiguities that are very different from the meaning-equivalent structures used in production research. In two experiments, we investigated whether priming during comprehension occurs in ditransitive sentences similar to those used in production research. When the verb was repeated between prime and target, we observed a priming effect similar to that in production. However, we observed no evidence for priming when the verbs were different. Thus, priming during comprehension occurs for very similar structures as priming during production, but in contrast to production, the priming effect is completely lexically dependent.
Keywords Language comprehension, Sentence processing, Syntactic priming, Anticipatory eye movements
2
It is generally assumed that recent exposure to a syntactic structure facilitates subsequent production and comprehension processes. Evidence that recent exposure to syntactic structures affects subsequent production comes from studies showing evidence for syntactic priming, the phenomenon that the processing of a syntactic structure facilitates the subsequent processing of the same structure (e.g., Bock, 1986, 1989; Bock & Loebell, 1990; Bock, Loebell & Morey, 1992; Pickering & Branigan, 1998). By contrast, it is much less clear whether syntactic priming affects comprehension processes in a similar way. Therefore, the question we address in the current paper is whether syntactic priming in comprehension occurs for similar structures as those tested in production research. One of the first studies showing evidence for syntactic priming during production was reported in a seminal paper by Bock (1986). Participants in her study produced ditransitive prime sentences that either had a prepositional object dative (PO) or double object dative structure (DO).
(1a) The lifeguard tossed the struggling child a rope. (DO structure) (1b) The lifeguard tossed a rope to the struggling child. (PO structure)
They were followed by a semantically unrelated target picture that could be described either with a PO (e.g., The man is reading a book to the boy) or DO structure (e.g., The man is reading the boy a book). The results showed that participants tended to describe the picture using the same syntactic structure as in the prime sentence. That is, after producing a PO prime (1b), they were more likely to describe a following picture using a PO structure than a DO structure, and likewise, after hearing a DO prime (1a), they were more likely to describe the picture using a DO structure than a PO structure. Subsequent experiments have suggested that sentence priming effects have a truly syntactic component and are not merely caused by non-syntactic factors. For example, Bock (1989) observed that priming occurred regardless of whether function words (e.g., to) were the same in prime and target. She also showed that 3
the priming effect was not due to prosodic similarities between prime and target. Bock and Loebell (1990) showed that sentences with the same structure as dative POs but with different semantic roles (e.g., The wealthy widow drove her Mercedes to the church) primed the production of PO structures to the same extent as dative POs, suggesting that the priming effect is syntactic rather than semantic in nature. Hartsuiker and Westenberg (2000) observed priming for different orderings of the past participle and auxiliary in Dutch (e.g., was stolen vs. stolen was). These structures do not differ in their conceptual or information structures, so these results suggest that sentence priming effects are not merely due to repetition of conceptual or information structure. Finally, Pickering and Branigan (1998) found that priming was unaffected by repetition of the verb's aspect, tense, or number in ditransitive structures such as (1). In sum, there is strong evidence for a syntactic component to sentence priming effects. In addition, there may also be other factors that play a role. For example, recent production experiments have provided evidence for semantic effects (e.g., Chang, Bock, & Goldberg, 2003; Cleland & Pickering, 2003; Griffin & Weinstein-Tull, 2003). Pickering and Branigan (1998; Branigan, Pickering, & Cleland, 2000) showed that repetition of the verb between prime and target is one important factor affecting syntactic priming. Using ditransitive structures such as (1), they showed that priming occurred when the verb in the prime and target was different, but it was much larger when the verb was the same. They explained these results with a model that assumes that syntactic structures are represented at a lexical level. In their model, each verb has separate connections to combinatorial nodes, which represent syntactic structures. When a particular verb occurs with, say, the PO structure in the prime, both the activation of the PO combinatorial node and the strength of the connection between the verb and the PO node are increased. Both result in priming when the verb in the prime and target are the same, so strong priming should occur. In contrast, when the verb in the target is different from the prime verb, the increased verbPO connection strength does not have any effect because it does not apply to the target verb,
4
so only the activation of the PO node results in priming. Hence, consistent with Pickering and Branigan's (1998) results, the priming effect should be weaker. It has also been shown that sentence production is affected by the comprehension of a preceding sentence, suggesting that sentence production and comprehension share common processing mechanisms or representations. Branigan et al. (2000) found that in dialogue, participants tended to repeat syntactic structures that were produced by their interlocutor (See also Cleland & Pickering, 2003). Drawing on this and other research, Pickering and Garrod (2004) proposed a model that assumes that interlocutors interactively align their utterances at all levels of representation in both production and comprehension. This predicts that syntactic priming should occur in comprehension as well as in production and suggests that the priming effects should be affected by similar factors. Most sentence comprehension theories do not make precise predictions about how exposure to a prime sentence should affect subsequent processing of the next sentence. However, nearly all theories assume that frequent exposure to a particular syntactic structure should have long-term effects resulting in facilitation of that structure during subsequent processing. This assumption explains why people find it easier to process syntactic structures that are highly frequent than structures that are infrequent (e.g., Clifton, Frazier, & Connine, 1984; Garnsey, Pearlmutter, Myers, & Lotocky, 1997; Mitchell & Holmes, 1985; Trueswell, 1996; Trueswell, Tanenhaus, & Kello, 1993). Although it is a controversial issue whether people immediately use syntactic frequency information during sentence processing (e.g., MacDonald, Pearlmutter & Seidenberg, 1994; Trueswell et al., 1993) or whether its use is delayed (e.g., Ferreira & Henderson, 1990; Frazier, 1995), nearly all sentence processing theories share the assumption that prior exposure to syntactic structures should have an effect at some stage during sentence processing. However, as highlighted by Mitchell, Cuetos, Corley, and Brysbaert (1995), an important question concerns the type of syntactic frequency information that people use. Mitchell et al. (1995) proposed that the processor only uses coarse-grained frequency information, that is, it does not take into account the frequency with 5
which syntactic structures co-occur with specific words. In contrast, constraint-based lexicalist theories claim that the sentence processor employs syntactic frequency information that is associated with individual lexical items (e.g., Garnsey et al., 1997; Juliano & Tanenhaus, 1994; MacDonald et al., 1994; Trueswell et al., 1993). In fact, most of these models assume that both lexically specific and lexically independent (coarse-grained) frequency information affect sentence processing (e.g., Juliano & Tanenhaus, 1994; McRae, Spivey-Knowlton, & Tanenhaus, 1998; Tabor, Juliano, & Tanenhaus, 1997). It is currently unclear to what extent long-term exposure and recent exposure in the previous sentence are similar, but what is clear is that syntactic priming effects during comprehension may also be either lexically specific or lexically independent in origin. The current study will investigate to what extent syntactic priming in comprehension is lexical in nature. This issue is highly relevant for the comparison between syntactic priming in comprehension and production, because, as demonstrated by Pickering and Branigan (1998), syntactic priming in production is partly lexically independent (resulting in priming when the verb in prime and target is not repeated) and partly lexically specific (resulting in stronger priming with repeated verbs). In order to facilitate the comparison between syntactic priming during production and comprehension, we employed ditransitive structures such as in (1), which are similar to those used by Pickering and Branigan (1998) and many other studies showing evidence for syntactic priming during production (e.g., Bock, 1986, 1989; Bock & Griffin, 2000; Bock & Loebell, 1990; Branigan et al., 2000; Branigan, Pickering, Stewart, & McLean, 2000; Branigan, Pickering, & Cleland, 1999; Corley & Scheepers, 2002; Pickering, Branigan, & McLean, 2002; Potter & Lombardi, 1998). Although syntactic priming during comprehension has received much less attention than syntactic priming during production, a few studies have investigated whether recent exposure affects subsequent comprehension of the same structure. Mehler and Carey (1967) found that participants perceived sentences masked by white noise significantly better if they were preceded by sets of sentences with a similar structure. For example, a sentence such as 6
They are describing events was perceived better following sentences such as They are forecasting cyclones than sentences such as They are recurring mistakes. Likewise, Mehler and Carey (1968) and Carey, Mehler, and Bever (1970) observed that participants' performance on a sentence-picture matching task was affected by prior exposure to similar sentences. However, in all these studies, a large number of prime sentences preceded the target trial, so it is likely that participants were aware of similarities between the primes and targets and that the results were due to strategic processes. Similarly, participants may have been aware of similarities between the sentences in a more recent study by Noppeney and Price (2004), who conducted a self-paced reading and fMRI experiment. Their materials consisted of temporarily ambiguous reduced relative sentences such as The child left by his parents played table football, where the child left can temporarily be analysed as a main clause (e.g., the child left his parents) and temporarily ambiguous intransitive structures such as After the headmaster had left the school deteriorated rapidly, where After the headmaster had left the school can temporarily be analysed as a transitive clause. Participants also received blocks of sentences that were either syntactically similar or dissimilar. Noppeney and Price observed that reading times were faster and brain activation was lower for blocks of sentences with syntactically similar structures. However, it is possible that this is due to strategic processes that occurred when participants discovered that the sentences in a block were syntactically similar. Branigan, Pickering, and McLean (2005) used a sentence-picture matching task in which target trials were preceded by a single prime sentence. In one of their experiments, participants first read globally ambiguous sentences such as (2).
(2) The policeman prodding the doctor with the gun.
The sentences were ambiguous in that the PP with the gun could be interpreted either as modifying the verb prod (VP attachment) or as modifying the noun phrase the doctor (NP 7
attachment). Participants next saw two pictures and were asked to choose the picture that they thought matched the prime sentence. In the prime trials, only one of the two pictures correctly matched the sentence, so the picture disambiguated the sentence. In contrast, in the target trials immediately following the prime trials, one of the pictures matched the VP attachment interpretation and the other the NP attachment interpretation. When the verb was repeated in prime and target, participants were more likely to choose the picture that matched the same interpretation as that in the prime trials, but there was no significant priming effect when the verb was not repeated. However, the priming effect may have been caused by perceptual similarities between the pictures in the prime and the target. The VP attachment pictures always showed an action being performed with an instrument (e.g., prodding with a gun) whereas the NP attachment pictures always showed an action without the instrument (e.g., prodding with a finger). Hence, the observed effect may not be linguistic. Furthermore, the sentence-picture matching task may be sensitive to strategic effects that are involved in decision processes, so it may not inform us about online sentence processing. Pickering and Traxler (2004) and Traxler and Pickering (2005) measured eye movements during reading to investigate syntactic priming. They employed sentences containing temporarily ambiguous reduced relatives such as in (3).
(3) The defendant examined by the lawyer turned out to be unreliable.
When the relative clause verb (examined) stayed the same between the prime and target, reduced relative sentences such as (3) were read faster when they were preceded by a sentence that was disambiguated towards the reduced relative interpretation than when they were preceded by a sentence disambiguated towards the main clause analysis. Furthermore, unreduced relative sentences (containing who was preceding the relative clause verb) and passive sentences also primed reduced relatives. However, no priming effects occurred when 8
the verb in the prime and target was different, suggesting that that priming in comprehension may be entirely dependent on verb repetition. In an ERP study, Ledoux, Traxler, and Swaab (2006) tested similar reduced relative materials and used the same verb in prime and target. Interestingly, they observed that the positive shift in the ERP wave signal (P600) was higher when a reduced relative was preceded by a main clause than when it was preceded by another reduced relative. Assuming that the P600 component reflects syntactic processing (e.g., Hagoort, Brown, & Groothusen, 1993; Osterhout & Holcomb, 1992), this suggests that the priming effects in reduced relatives are syntactic rather than semantic in nature. In all the above studies, the structures that were tested were very different from those used in production studies, so it is difficult to draw comparisons of syntactic priming effects in production and comprehension. Furthermore, the prime sentences in the different conditions differed in syntax as well as in semantics. Except for the ERP study by Ledoux et al. (2006), where the P600 component suggests that the priming effect is syntactic, the effects in the other studies may be due to semantic rather than syntactic properties of the primes. Frazier, Taft, Roeper, Clifton, and Ehrlich (1984) did use sentences that were equivalent in meaning. They investigated parallelism effects for a variety of different structures, including active and passive sentences that have been tested in production studies of syntactic priming (e.g., Bock, 1986, 1989; Bock & Loebell, 1990). They observed that reading times for the second clause in sentences consisting of two active clausal conjuncts (4a) were faster than in sentences consisting of a passive and active conjunct (4b).
(4a) The tall gangster hit John and the short thug hit Sam. (4b) John was hit by the tall gangster and the short thug hit Sam.
This may suggest that the second conjunct was syntactically primed by the first. However, Frazier, Munn, and Clifton (2000) observed that parallelism effects occurred with conjoined 9
phrases such as a strange man and a tall woman in Hilda noticed a strange man and a tall woman when she entered the house, but not with similar non-conjoined phrases as in a strange man noticed a tall woman yesterday at Judi’s, suggesting that parallelism effects are not due to genuine syntactic priming. Frazier et al. suggested that parallelism effects in conjoined phrases may occur because the structure of two conjoined phrases is often the same, so the first phrase predicts the structure of the second phrase. By contrast, in non-conjoined phrases, the structure of the first phrase does not reliably predict the structure of the second. In addition, we believe that parallelism effects may also be due to stylistic preferences for parallel conjoined structures. Furthermore, in many of Frazier et al.’s (1984) materials, lexical items and semantic content were repeated across the two conjuncts, so it is unclear whether their parallelism effects occur in the absence of such repetition. Branigan (1995) showed no evidence for parallelism effects in active/passive structures and PO/DO structures with materials that did not have lexical and semantic repetitions, whereas there were significant parallelism effects for temporarily ambiguous sentences such as reduced relatives. More recently, Luka and Barsalou (2005) also investigated priming from meaningequivalent sentences. In their grammaticality acceptability experiment, they tested ditransitive PO/DO sentences such as (5) as part of a larger set of sentence structures.
(5a) Ramarez passed Santiago the ball. (5b) Ramarez passed the ball to Santiago.
They found that participants judged a sentence as more acceptable if it was preceded by sets of structurally similar prime sentences. However, they did not separately analyse priming effects for the different types of sentence structures they used, so it is not possible to ascertain whether syntactic priming occurred for PO/DO structures such as (5). Furthermore, acceptability judgements reflect conscious decisions that are likely to be affected by 10
relatively late and strategic processes and they encourage participants to pay particular attention to structural features of sentences. Hence, it is unclear what the implications for normal online sentence processing are. Scheepers and Crocker (2004) used the visual-world eye-movement paradigm to investigate syntactic priming, a method known to be highly sensitive to moment-to-moment sentence processing effects (e.g., Altmann & Kamide, 1999; Eberhard, Spivey-Knowlton, Sedivy, & Tanenhaus, 1995; Tanenhaus, Spivey-Knowlton, Eberhard, & Sedivy, 1995; Trueswell, Sekerina, Hill, & Logrip, 1999). In Scheepers and Crocker's experiment, participants first read aloud written German prime sentences such as (6a) or (6b). The constituent order in the prime sentences was either subject-verb-object (SVO, 6a), which is the canonical order in German, or object-verb-subject (OVS, 6b).
(6a)
Der Regisseur lobte insbesondere den Produzenten. The director [nom] commended in particular the producer [acc].
(6b)
Den Regisseur lobte insbesondere der Produzent. The director [acc] commended in particular the producer [nom].
(7a)
Die Krankenschwester föhnt offensichtlich den Priester. The nurse [ambig.] blow-dries apparently the priest [acc].
(7b)
Die Krankenschwester schubst offensichtlich der Sportler. The nurse [ambig.] pushes apparently the sportsman [nom].
The prime sentences (6) were unambiguous by virtue of the case marking on the first article (der/den). Participants next listened to an auditorily presented target sentence such as (7a) or (7b), which was temporarily ambiguous, because case marking on the first article (die) permitted both a subject and object reading of the first NP. At the same time, participants saw a picture showing two events. In one of the events, the first NP was an agent and the 11
second a patient (e.g., the nurse blow drying a priest), so this event was consistent with a subject interpretation of the first NP (and therefore with the SVO analysis). In the other event, the first NP was a patient and the second an agent (e.g., the nurse being pushed by a sportsman), consistent with an object interpretation of the first NP. Scheepers and Crocker observed that the constituent order in the prime affected participants' anticipatory eye movements to the second noun (priest/sportsman) while they listened to the case ambiguous first NP (e.g., die Krankenschwester). After SVO primes, participants made more looks to the character that was depicted as being a patient (e.g., the priest) than to the character depicted as an agent (e.g., the sportsman), whereas it was the other way around after OVS primes. Hence, participants anticipated the second noun on the basis of the structure of the prime in combination with the information in the visual scene. The current study also used the visual-world eye-movement paradigm to investigate syntactic priming. In order to gain insight into similarities between syntactic priming in comprehension and production, we tested ditransitive PO/DO structures that have previously been used to study sentence production (e.g., Bock, 1986, 1989; Pickering & Branigan, 1998). Furthermore, unlike the materials in many previous studies investigating syntactic priming in comprehension, the two structural variants of the ditransitive construction (PO or DO) have essentially the same meaning, so any observed priming effects must be due to the structure (i.e. constituent order) rather than due to the semantics of the primes. Participants first read a PO or DO prime sentence such as (8) aloud.
(8a) The assassin will send the dictator the parcel. (8b) The assassin will send the parcel to the dictator.
(9a) The pirate will send the princess the necklace. (9b) The pirate will send the necklace to the princess.
12
Next, they listened to a PO or DO target structure such as (9) while their eye movements to a visual scene containing a picture of the agent (the pirate), recipient (the princess) and theme (the necklace) in the sentence were monitored. Using the visual-world paradigm, Altmann and Kamide (1999) showed that people anticipate upcoming linguistic information. In their study, participants listened to sentences such as the boy will eat the cake while they saw a visual scene containing a boy, a cake, and a number of distractor objects. As soon as participants heard the verb eat, they tended to look at the cake, suggesting that they anticipated properties of the arguments following the verb. Subsequent experiments (Kamide, Altmann, & Haywood, 2003; Kamide, Scheepers, & Altmann, 2003) have shown that combinatorial semantic information provided by two words in the sentence and case marking in languages such as German and Japanese also influence anticipatory eye movements. On the basis of these earlier findings, we predict that when participants hear the verb and anticipate a recipient role immediately following it, they should look more often and longer at the recipient picture than when they anticipate a theme role. Similarly, they should look more often and longer at the theme picture if they anticipate a theme role immediately following the verb than if they anticipate a recipient role. Hence, if syntactic priming occurs and influences anticipatory eye movements, we expect that participants will look more often and longer at recipient pictures following DO primes than following PO primes, and look more often and longer at theme pictures following PO primes than following DO primes. Apart from demonstrating that syntactic priming occurs during comprehension with similar structures to those used in production studies, such results would also have important implications regarding the nature of the anticipations that people make. It is currently unclear whether anticipations are purely semantic or include a more structural component. In the studies by Altmann, Kamide, and colleagues, participants may have anticipated semantic properties of all possible arguments following the verb without regard to the order in which the arguments should appear (a structural component). If the priming manipulation in our 13
study affects anticipatory eye movements, this would suggest that participants anticipate information related to the order of the arguments following the verb. Such a finding would be difficult to reconcile with models that assume that the processor does not project any structure until it encounters the noun phrase following the verb (e.g., Frazier, 1987). Instead, it would support the view that the processor projects syntactic information at the verb (e.g., Gorrell, 1995; MacDonald et al., 1994; Pritchett, 1992). This article reports two experiments. In Experiment 1, we investigated whether syntactic priming occurs when the verb in the prime and target sentence is the same, as in (89). In Experiment 2, the verb in the prime and target was different. Comparison of the results from the two experiments enabled us to test whether syntactic priming in ditransitive sentences during comprehension is lexically specific or independent. We will return to this issue in Experiment 2.
Experiment 1
Experiment 1 investigated syntactic priming in comprehension in PO/DO structures that had the same verb in prime and target.
Method
Participants. Thirty-two participants from the University of Dundee student community were recruited to take part in the experiment. All were native speakers of British English with normal or corrected-to-normal vision. They received either course credit or £4.00 in exchange for their participation.
Materials. Thirty-two items were constructed, each consisting of two stimuli: a written sentence serving as the prime, and a spoken sentence combined with a semi-realistic scene 14
serving as the target. Prime sentences were either of the double object form (DO, e.g., 8a) or of the alternative prepositional object form (PO, e.g., 8b). The two types of prime sentences differed in the syntactic structure of the verb phrase (VP). In DO sentences, the verb was followed by two noun phrases (NPs): an indirect object denoting the recipient and a direct object referring to the theme. In PO sentences, the verb was followed by a direct object NP denoting the theme and a prepositional phrase (PP) denoting the recipient. The spoken target sentences were also either in the DO (9a) or PO (9b) form, like the written prime sentences they were paired with. Target sentences were recorded on a MiniDisk from a male native speaker of Scottish English who was instructed to use a neutral intonation. The recordings were saved as 16 KHz mono wave-files for later editing and presentation. The form of the target sentences was manipulated orthogonally to the form of the prime sentences, making it impossible to predict the eventual target structure from the type of prime preceding it. Except for the verb, none of the content words was repeated between prime and target, and the prime and target sentence could not plausibly be interpreted as part of the same discourse. Finally, in all of our target sentences, both referent nouns following the verb started with a consonant (pronunciation of the postverbal determiner the was therefore not indicative of the following noun). The 32 sets of prime and target sentences were generated from a list of 12 ditransitive verbs (see Appendix). These were taken from previous studies on language production and were known to support PO and DO constructions about equally well. Each target sentence was combined with a semi-realistic visual scene, created from commercially available clipart images. They were saved as 16-bit colour bitmaps in 1024×768 pixels resolution. Each picture contained three entities, each of which corresponded to one of the thematic roles (agent, recipient, and theme) in the corresponding ditransitive target sentence. Figure 1 provides an example.
[Insert Figure 1 here] 15
The pictured entities were designed to be easily recognisable and had no obvious semantic relation to one another. The three entities in each picture were arranged in a ‘triangular’ fashion: one entity was placed on the left, one entity in the middle (above or below centre), and one entity on the right of the display. Across items, we ensured that agent, theme, and recipient entities appeared equally often in each of the three positions, thereby counterbalancing potential scanning preferences. Pictures always stayed the same across experimental conditions – only the linguistic materials (prime and target sentences) varied according to the experimental design.1
Design and Procedure. We employed essentially the same methodology as in Scheepers and Crocker (2004), combining reading of the prime sentences with a visual-world paradigm in the target trials. For each picture (e.g., Figure 1), each type of prime (8a vs. 8b) was combined with each type of target sentence (9a vs. 9b), resulting in a 2 × 2 repeated measures design. The experiment contained 32 experimental items, each having four conditions. We constructed four lists comprising eight items from each condition, with exactly one version of each item appearing in each list. Additionally, 35 written sentences and 39 visual scenes plus auditory sentences were included as fillers in each list. Filler sentences were syntactically and semantically unrelated to the ditransitive structures under investigation and consisted of intransitives, passives, or copula-verb constructions. Consistent with the experimental target items, the visual-world fillers employed pictures composed of three entities in a triangular arrangement. However, in the visual-world fillers, it was often the case that the auditory sentence did not refer to all the entities in the picture (e.g., Cleopatra’s favourite food is icecream presented in combination with a picture showing Cleopatra, a cone of ice-cream, and Frankenstein’s monster). The items of each list were presented in a fixed quasi-random order, subject to the constraint that the critical prime-target pairs were preceded by at least two filler items 16
randomly chosen from the full set of reading and visual-world fillers. This ensured that the pairings of primes and targets were not transparent to the participant (i.e. it was not the case that a reading trial was always followed by a visual-world trial, or vice versa). Four fillers appeared at the beginning of each experimental session and three fillers followed a short break half way through the experiment. Participants were seated approximately 75 cm from a 21” colour monitor running at 120 Hz. Throughout the entire experiment, they wore an SR-Research EyeLink-II headmounted eye-tracking system with a sampling rate of 500 Hz and a spatial resolution of less than 0.01°. Viewing was binocular, but only the participant’s dominant eye was tracked (the right eye for about 70% of the participants, as determined by a simple parallax test prior to the experiment). Head movements were not restricted. However, participants were instructed to keep head movements to a minimum during the experiment. Auditory stimuli were presented via a satellite-speaker and subwoofer system. The eye-tracker continuously recorded onsets and offsets of fixations (as defined by acceleration and velocity thresholds) together with corresponding spatial coordinates (in pixels). The experiment began with the adjustment of the IR cameras of the tracker, a procedure that took about one minute per participant. Next, a brief calibration procedure was performed during which the participant had to look at a fixation cross in nine different positions on the screen. This procedure, which took about 30 seconds, was repeated after a short break half way through the experiment and whenever measurement accuracy appeared insufficient (e.g., after changes in the participant’s posture). However, re-calibration was never performed between a critical prime-target pair of trials. A reading trial always started with the presentation of a fixation cross in the centre of the screen. The participant had to fixate this cross such that an automatic drift-correction could be performed. The experimenter pressed a button to trigger the presentation of the written sentence. Participants were instructed to read the written sentences aloud, and to press a button to proceed to the next trial when they had finished reading. 17
A visual-world trial always appeared immediately after the previous trial had ended, that is, without prior presentation of a fixation cross. This ensured that the time between a prime trial and the following target trial was kept to a minimum (200 ms average ISI). Participants were informed that the spoken sentences would always refer to the visual scenes they were combined with, and that it was important to pay attention both to the visual scene and the spoken sentence. Auditory sentence presentation always started after a 1000 ms ‘preview period’ following the onset of the picture. The picture stayed on the screen for seven seconds (roughly one second longer than the end of the sentence), after which the next trial was initiated. Participants were told that that the experiment was concerned with both written and auditory language comprehension and were unaware of the actual purpose of the study until debriefing at the close of the experiment. To keep participants focused, six visual-world fillers were followed by a written prompt indicating that participants had to verbally describe the previous scene in their own words (participants typically responded by repeating the auditory sentence they had heard before). Each experimental session took about 30 - 40 minutes to complete.
Data Analysis and Results
The spatial fixation coordinates from the original tracker output were mapped onto four different areas in the visual scene: agent, recipient, theme, and background. This was done using analysis bitmaps, which were graphically edited versions of the original presentation bitmaps. The three objects (agent, recipient, and theme) in each picture were manually defined, allowing an area of approximately 30 pixels around the contour of an object to be coded as belonging to that object. Areas that were not associated with an object in this way were coded as background. Our analysis software used these areas to identify whether participants fixated on the picture of the recipient, theme, agent, or background. The same 18
software also pooled extremely short fixations (with a duration of less than 80 ms) with immediately preceding or following fixations if those fixations lay within 0.5 degrees of visual angle (ca. 12 pixels) from the short fixations (otherwise short fixations were eliminated from analyses). The rationale behind this procedure was that such short fixations often result from false saccade planning and are unlikely to reflect visual information processing (e.g. Rayner & Pollatsek, 1989). To enable synchronisation of eye movements with corresponding events in the auditory sentences, the onset of the main verb, the postverbal determiner, and the first postverbal noun in each target sentence were hand-coded in millisecond-resolution using professional sound editing software. On average, the onset of the main verb occurred 2364 ms (DO targets) and 2338 ms (PO targets) after picture onset. The difference was not significant (t (31) = 1.22, p = .23) according to a within-items t-test. The onset of the postverbal determiner occurred 3328 ms (DO targets) and 3312 ms (PO targets) after picture onset (t (31) = .60, p = .55), and for the first postverbal noun, the figures were 3479 ms (DO targets) vs. 3462 ms (PO targets); again, the difference was not significant (t (31) = .65, p = .52). Of particular interest in all our analyses were fixations on the recipient and theme objects in the target trials following the onset of the verb. Assuming that participants are able to anticipate the ordering of the postverbal arguments at the verb and that this is reflected in anticipatory eye movements, we expected that anticipatory looks to the recipient object before the presentation of the first postverbal noun would indicate that perceivers are expecting a DO construction to follow (a recipient NP immediately follows the verb in this structure); conversely, anticipatory looks to the theme object would indicate that listeners are expecting a PO construction (where a theme NP immediately follows the verb). Consequently, if perceivers are primed by previous exposure to one of the two constructions, we would expect that after having read a DO prime, participants should deploy more visual attention to the recipient rather than the theme object when they hear the verb in the target 19
trial; after having read a PO prime, however, they should be more likely to focus attention on the theme rather than the recipient object when hearing the verb.
First-Gaze Durations. In the first set of analyses, we focused on average first-gaze durations to the recipient and theme objects. Following standard definitions, a gaze was defined as the accumulation of all consecutive fixations on a picture object until another object (or the background) was fixated. We analysed average gaze durations that started between the onset of the verb and the onset of the postverbal determiner (gazes that were initiated before the verb or after the onset of the postverbal determiner were ignored). Moreover, we only included first gazes initiated between the onset of the verb and the postverbal determiner. That is, if people fixated on, for example, the recipient, the theme and the recipient (in that order) after the onset of the verb, we only analysed the very first gaze on the recipient, while discarding the gaze on the theme and the second gaze on the recipient. Table 1 shows the mean first-gaze durations on the recipient object and the theme object separately for each prime condition. Note that we collapsed across target structure (DO vs. PO), because both types of target sentences were identical up to the onset of the postverbal noun. Indeed, there were no main effects or interactions involving target structure (Fs < 1) in the first-gaze durations, justifying our approach of simply collapsing across levels of this variable.
[Insert Table 1 here]
Data analyses were conducted using mixed-design ANOVAs including prime type (DO vs. PO) and picture object (recipient vs. theme) as within-participants (F1) and withinitems (F2) variables; the four-level variable participant/item list was included as an additional between-participants or -items variable in order to eliminate random variance between groups (Pollatsek & Well, 1995). The analyses revealed no main effects of either prime type (Fs < 20
1) or picture object (Fs < 1). However, the interaction between the two variables was significant (F1 (1, 28) = 11.92, p < .01; F2 (1, 28) = 6.00, p < .05), indicating a priming effect. Simple effect comparisons indicated that first gazes on recipient objects were longer after DO than PO primes, although this effect was marginal by participants (F1 (1, 28) = 3.43, p = .08; F2 (1, 28) = 5.35, p < .05). Conversely, first gazes on theme objects were longer after PO than DO primes, although this effect was not significant by items (F1 (1, 28) = 10.87, p < .01; F2 (1, 28) = 1.43, p = .24). The simple effects suggest that priming occurred for both PO and DO targets, but they did not reach standard levels of significance by both participants and items because the priming effect on each target separately was smaller than the overall priming effect. Most crucially, given the significant interaction, we conclude that durations of first gazes launched between verb and postverbal determiner onset provided evidence for priming.
Log gaze probability ratios. In addition to first-gaze durations, we analysed gaze probabilities per picture object over time. Because proportions of looks to different picture objects are not linearly independent of one another (more looks to object A imply fewer looks to object B, and vice versa), we did not include picture object as a variable in the ANOVAs. Instead, we defined an analysis measure that expresses the strength of the visual bias towards the recipient object relative to the theme object in terms of a log-ratio: ln(P(recipient)/P(theme)), where P(recipient) refers to the likelihood of gaze on the recipient object, and P(theme) the likelihood of a gaze on the theme object. The measure yields a score of zero if both picture objects are inspected equally frequently (e.g., ln(0.2/0.2) = ln(1) = 0); a positive score means that there are more looks to the recipient than the theme object (for instance, a 3:2 advantage for the recipient object yields ln(0.3/0.2) = ln(1.5) = 0.405), and a negative score implies that there are more looks to the theme than the recipient object (e.g., a 3:2 advantage for the theme object translates into ln(0.2/0.3) = ln(0.667) = −0.405). Thus, the measure is symmetrical around zero. Its absolute score reflects the magnitude, while its sign 21
expresses the direction of the visual bias. Another advantage of log-ratios over standard probabilities is that they yield data distributions that are more suitable for parametric testing (standard probabilities often imply a violation of the homogeneity of variance assumption because they can only take values between 0 and 1; log-ratios, by contrast, can take values between minus infinite and plus infinite). Note that error term dfs in the log-ratio analyses varied due to case-wise exclusion of participants or items that yielded zero gaze probabilities on one or the other picture object during the relevant time window (the log of zero is not defined, as is division by zero)2. Analyses relative to verb onset. First, we conducted analyses of the log gaze probability ratios relative to the onset of the verb. Starting from the verb onset in each individual target sentence, we divided the gaze record into time slices of 20 ms each. Figure 2 plots the mean log gaze probability ratios (as defined above) over 75 of these time slices from the onset of the main verb, spanning a period of 1500 ms. For inference-statistical analysis, these 20 ms time slices were aggregated into larger time windows. That is, we divided the 1500-ms time window following the onset of the main verb into five 300-ms bins and then calculated the log gaze probability ratio for each bin.
[Insert Figure 2 here]
We conducted analyses of the logprobability ratio data within each 300 ms time window following the onset of the verb using mixed-design ANOVAs including prime type as withinparticipants (F1) and -items (F2) variables, and participant/item list as a between-participants and -items variable. Both in the analyses and in Figure 2, we collapsed across the variable target structure (PO vs. DO) because our analyses revealed no effect of this variable in the first 1500 ms after the verb onset (Fs < 1). Within 0-300 ms from the verb onset, Figure 2 shows a clear positive trend in the logratio data, indicating an overall bias towards the recipient object in this time window. This 22
bias was confirmed by the intercept estimate in the ANOVA, which differed reliably from zero (F1 (1, 26) = 10.88, p < .01; F2 (1, 27) = 4.36, p < .05)3. This bias may be due to a preference for animate objects (recipients were always animate while theme objects were always inanimate) or due to differences in visual saliency or size between the recipient and theme picture objects. There was no effect of prime type during the 0-300 ms time window (Fs < 1), which is unsurprising given that this time window is too early for participants to process the verb. In the following 300-600 ms time window, we found the same overall recipient bias in the intercept statistics (F1 (1, 27) = 10.72, p < .01; F2 (1, 28) = 7.70, p = .01), again suggesting a preference for animate objects. The effect of prime type was not significant during this time window (F1 < 1; F2 (1, 28) = 2.31, p = .14). The 600-900 ms time window after verb onset revealed no overall bias towards one or the other picture object (Fs < 1 for the intercept). Instead, there was a significant effect of prime type (F1 (1, 28) = 4.42, p < .05; F2 (1, 28) = 5.64, p < .05), indicating that participants’ looks to the recipient and theme were affected by whether they had just read a PO or DO sentence aloud. Analyses for the 900-1200 ms time window after verb onset also showed no overall visual bias (Fs < 1 for the intercept), but again a clear effect of prime type (F1 (1, 28) = 8.29, p
