Ординатура / Офтальмология / Английские материалы / Eye Movements A Window on Mind and Brain_Van Gompel_2007

Abstract

Compound words present an opportunity to study the organization of the mental lexicon. Decomposition of compound words occurs in multiple languages. Most studies of compound words have used semantically transparent compounds. It is therefore difficult to localize the level at which decomposition occurs. The present study explored the role of semantic transparency for English compound words. Transparent and opaque compounds were embedded into sentences and the frequencies of their lexemes were manipulated. Analysis of gaze durations revealed main effects of lexeme frequency and transparency. Transparency did not interact with lexeme frequency, suggesting decomposition occurs for both transparent and opaque compounds.

first fixation durations as well as gaze durations (a measure that takes refixations on a word into account) on the compound. Pollatsek, Hyönä, and Bertram (2000) also observed effects of second lexeme frequency and whole compound frequency in gaze durations. Based on these results, Pollatsek et al. proposed a dual-route theory of compound word recognition, according to which there is both a direct route to compound word meaning as well as a route through the decomposed lexemes.

Recent eye-movement experiments with English compound words have also found evidence of morphological decomposition. Juhasz, Starr, Inhoff, and Placke (2003) orthogonally manipulated the frequencies of beginning and ending lexemes in familiar compound words. A small, non-significant effect of beginning lexeme frequency was observed on first fixation durations as well as a larger, more robust effect of ending lexeme frequency in gaze duration. Based on these findings, Juhasz et al. concluded that the ending lexeme may have a privileged role in the recognition of English compound words.

Andrews, Miller, and Rayner (2004) also examined the processing of English compound words in an eye-movement study where lexeme frequencies were manipulated. Unlike the Juhasz et al. study, they observed large effects of beginning lexeme frequency in gaze duration and a smaller effect of ending lexeme frequency. Similar to the Finnish data, Andrews et al. also observed a whole-word frequency effect for compound words in a regression analysis. Thus, these results support the hypothesis that compounds are decomposed during recognition, but also have a whole-word representation. It appears that some of the difference between the Juhasz et al. findings and Andrews et al. findings in terms of importance of lexeme location may be due to the length of the compound words used in the two studies. In the Juhasz et al. study, all compound words were 9 letters in length. In the Andrews et al. study, compound word lengths varied between 6 and 11 characters. Recent results by Bertram and Hyönä (2003) have suggested less of an influence for the beginning lexeme frequency for shorter compound words (7–9 letters) in Finnish. Thus, the addition of longer compounds in the Andrews et al. study may be one of the reasons why larger beginning lexeme effects were observed compared to the Juhasz et al. study.

The clear result from the above studies is that compound words are decomposed at some point during their recognition. In addition, compound words appear to be represented as a whole word in the lexicon as well. What the studies cannot speak to is how the decomposed lexemes are being used to access the compound word representation. In the studies mentioned above, most of the compound words were semantically transparent. A semantically transparent compound word is one in which both lexemes contribute to the meaning of the compound word (e.g., farmhouse). This can be contrasted with a semantically opaque compound word where the lexemes do not contribute to the meaning of the compound (e.g., snapdragon). For semantically transparent compound words, the lexemes in the compound word are related to the compound word on both a morphological level as well as a semantic level.

In one influential view of morphology (Marslen-Wilson, Tyler, Waksler, & Older, 1994), the role of morphemes is to provide a relationship between word forms and word

meanings. According to this view, only semantically transparent complex words are connected to the representations of their morphological constituents. However, recent work using the masked priming technique has called this view into question. In this technique, a quickly presented prime word is masked. After the prime word appears a target word is presented for some type of response. If the duration of the prime word is short enough, participants are not consciously aware of it. Recently, Rastle, Davis, and New (2004) demonstrated significant priming using this technique for words with semantically transparent morphological relationships (cleaner/clean), and words with an apparent morphological relationship but no semantic relationship (corner/corn), but no priming for words consisting of an embedded word and an illegal suffix (brothel/ broth). These results suggest a very early morphological decomposition process that operates at a level in the lexicon prior to semantics (see also Christianson, Johnson, & Rayner, 2005).

There is also evidence using compound word stimuli for an early non-semantically mediated decomposition process. The majority of work investigating the role of semantic transparency for compound words has used priming in a lexical decision task. In this task, participants are provided with a string of letters and they must decide whether these letters make up a word. Monsell (1985) presented participants with English transparent compound words (e.g., tightrope), opaque compound words (e.g., butterfly), or pseudocompounds (e.g., furlong). In the prime phase, constituents of the three types of stimuli were presented for lexical decision. Following this, the actual compound words were presented for lexical decision. There was significant constituent repetition priming for transparent compounds, opaque compounds, and pseudocompounds in this experiment, suggesting that all compound words were decomposed at some level.

In contrast to Monsell (1985), Sandra (1990) did not find equivalent priming for transparent and opaque compound words. In a series of lexical decision experiments using Dutch stimuli, Sandra failed to obtain priming for the lexemes in opaque compounds and pseudocompounds but did obtain significant priming for transparent compound words. However, Sandra (1990) used a semantic prime for the constituents in the compounds (e.g., bread priming butterfly), as opposed to a repetition prime. The importance of this difference is highlighted by Zwitserlood (1994), who investigated the role of semantic transparency for German compound words. In one experiment, Zwitserlood used compound words that were transparent, partially opaque (where one constituent was related in meaning to the compound and the other one was not), or fully opaque (where neither constituent of the compound was related in meaning to the compound expression). These compounds served as primes for their first or second constituents in a lexical decision task. Significant priming was observed for all types of compounds in this experiment. In a second experiment, transparent, partially opaque, fully opaque, and pseudocompounds served as primes for a word semantically associated with one of their constituents. There was no priming of semantic associates for fully opaque compounds or pseudocompounds. However, significant semantic priming occurred for the transparent compounds and the partially opaque compounds. Based on these findings, Zwitserlood argued that all compound words (and pseudocompound words) are represented as morphologically complex

at some level in the lexicon, and that this information is represented at a stage prior to semantics.

Jarema, Busson, Nikolova, Tsapkini, and Libben (1999) also found significant constituent repetition priming in a lexical decision task for fully transparent, partially opaque, and fully opaque French compound words. For Bulgarian, significant constituent repetition priming was observed for transparent and partially opaque compounds, but not for fully opaque compounds. Using English compounds, Libben, Gibson, Yoon, and Sandra (2003) found significant constituent priming for transparent, partially opaque, and fully opaque compound words.

These studies support a fast decomposition process that is not tied to semantics. However, all of the studies have used the lexical decision task. This task is sensitive to other factors, such as the makeup of the stimulus list (see Andrews, 1986). The recording of eye movements while reading compound words provides a more natural way to study the role of semantic transparency in morphological decomposition. To date, there are only a small set of eye-movement studies examining the semantic transparency of compound words. Underwood, Petley, and Clews (1990) embedded semantically transparent and opaque compound words in sentences. They observed a significant effect of transparency on gaze durations, with longer gazes in the case of opaque compounds. These results can be contrasted with those of Pollatsek and Hyönä (2005), who manipulated the frequency of the first lexeme in Finnish bilexemic transparent and opaque compound words. They observed a significant effect of beginning lexeme frequency, but no significant effects of semantic transparency on gaze duration. Semantic transparency of the compound also did not interact with the frequency of the beginning lexeme. Thus, both semantically transparent and semantically opaque compound words were decomposed, supporting an early decomposition process.

In the present experiment, semantically transparent and semantically opaque English compound words were embedded in neutral sentence contexts. As mentioned above, two published eye-movement studies examining semantic transparency have yielded contradictory results as to whether transparency influences gaze durations. Therefore, one purpose of the present experiment was to test the hypothesis that semantic transparency influences gaze durations. In addition, the frequencies of the first and second lexemes were orthogonally manipulated to index morphological decomposition. This extends the work of Pollatsek and Hyönä (2005), since in that study only the frequency of the beginning lexeme was manipulated. Finally, this study also explored an additional measure compared to previous studies, go-past durations. This measure is particularly suited to examine how words are integrated with the beginning of the sentence. The hypothesis currently under examination is that both transparent and opaque compound words are decomposed initially due to a fast initial decomposition process that is not tied to the meaning of the compound word. When words are integrated into the sentence, the meaning of the word becomes particularly important. Thus, it is possible that different effects will be observed in a measure indexing sentence integration compared to initial word recognition.

2. Method

2.1. Participants

Twenty-four University of Massachusetts community members participated in exchange for extra course credit or eight dollars. All participants were native speakers of American English and had normal vision or wore contacts.

2.2. Apparatus

Eye movements were recorded via a Fourward Technologies Dual-Purkinje eye-tracker (Generation V) interfaced with an IBM compatible computer. Viewing was binocular; however, eye movements were only recorded from the right eye. The eye tracker has a resolution of less than 10 min. of arc. The sentences were displayed on a 15-inch NEC MultiSync 4FG monitor. The monitor was set 61 cm from the participant. At this distance, 3.8 characters equal one degree of visual angle.

2.3. Procedure

Participants were given instructions detailing the procedure upon arrival to the experiment. A bite bar was prepared and head rests were also used to stabilize the head. Prior to starting the experiment the eye-tracker was calibrated. This calibration was checked after each sentence throughout the experiment, and redone as necessary. Comprehension questions were asked on 10–15% of the trials by providing participants with a question that they could verbally answer “yes” or “no”.

2.4. Materials

Forty transparent and 40 opaque English bilexemic compound words were selected as the stimuli. Opaque and transparent compounds were initially chosen based on experimenter intuitions, with transparent compounds classified as those where both lexemes in the compound contribute to the overall meaning of the compound (e.g., dollhouse) and opaque compounds classified as compounds where the meaning of the compound word was not easily computable from the meaning of the two lexemes (e.g., pineapple). These intuitions were checked in an experiment where eight University of Massachusetts community members who did not take part in the eye-tracking experiment were asked to rate the compound words on a 1–7 scale in terms of how transparent the meaning of the compound words were, with higher numbers signaling greater transparency of meaning. There was a clear dissociation between the transparent and opaque compounds. On average, opaque compounds received a rating of 2.86 while transparent compounds received an average rating of 5.79. This difference between opaque and transparent compounds was significant (t 78 = −21 74, p < 001). Average transparency ratings did not differ as a function of lexeme frequency (t’s <1).

In addition to semantic transparency, the frequencies of the first and second lexeme were also orthogonally manipulated, resulting in four conditions for each transparency class (eight conditions in total). These conditions consisted of words where both lexemes were high in frequency (HH), where the first lexeme was high in frequency and the second lexeme was low in frequency (HL), where the first lexeme was low in frequency and the second was high in frequency (LH), and where both lexemes were low in frequency (LL).2 Lexeme frequency was calculated as each lexeme’s frequency of occurrence as an individual word. Frequencies were calculated from the CELEX English database written frequencies and scaled to be out of 1 million (Baayen, Piepenbrock, & Gulikers, 1995). High-frequency lexemes had a frequency of greater than 53 per million (range 53.18–1009.55) while low-frequency lexemes had a frequency of less than 39 per million (range 0–38.21 per million). Beginning lexeme frequency did not vary significantly as a function of ending lexeme frequency or transparency (t’s < 1). Ending lexeme frequency did not vary significantly as a function of beginning lexeme frequency or transparency (t’s < 1). Compound words were also controlled on overall word frequency, which did not vary significantly as a function of transparency or lexeme frequency (p’s > 1). Compound words ranged from 8 to 11 characters. Length of the compound words, and the length of their lexemes did not vary significantly as a function of transparency or lexeme frequencies (t’s < 1). In addition, the number of unspaced compounds occurring in the CELEX database (Baayen et al., 1995) with the same beginning lexeme was also calculated for each compound word to get a measure of morphological family size for the beginning lexeme.3 The average number of compounds containing the same beginning lexeme was 8.29 (range 1–106). As would be expected, this number was significantly greater for compounds containing a high-frequency lexeme (t 78 = 2 72, p < 01). Morphological productivity did not significantly vary as a function of ending lexeme frequency or transparency (t’s < 1). Table 1 provides information about the conditions and Table 2 provides examples of the materials.

Compound words were embedded in neutral sentence frames, with the condition that the compound could not occupy the first two or last two positions in the sentence. Each compound word was fit into its own neutral sentence frame. Each target compound was preceded and followed by a word of 5–8 characters. The pre-target word was carefully selected to be of a mid-range frequency (ranging from 21 to 72 words per million in the CELEX database, Baayen et al., 1995), while the post-target word frequency was allowed to vary (ranging from 11 to 180 words per million). Importantly, the frequency of the pre-target or post-target word did not significantly differ with respect to the frequency of lexemes, or semantic transparency of the compounds (all p’s > 30). Ten University of

2 Initially there were 10 items per condition. However, two items were incorrectly classified as LH transparent compound words. The items (battlefield and minefield) should actually have been classified as HH transparent compound words. Prior to analyzing the data, these items were reclassified, resulting in 12 items in the HH transparent condition and 8 items in the HL transparent condition. Importantly, the pattern of significant results did not change when these items were reclassified.

3 Plurals and derivations of the compounds that occurred in the database were not counted as separate compounds.

Table 1

Stimulus characteristics as a function of compound type (opaque or transparent) and the frequency of the beginning and ending lexeme

Note: Lex1 = beginning lexeme, Lex2 = Ending lexeme, Freq = average Celex written frequency per million (Baayen et al., 1995). Lengths are expressed in average number of letters per condition.

Massachusetts community members performed a cloze task where they were presented with the sentence up to the pre-target word and were asked to provide the next word in the sentence. Target predictability was very low, amounting to 2.5% or less on average for each condition. In addition, cloze task performance did not significantly differ as a function of transparency or lexeme frequency (all p’s > 12). All sentences were less than 80 characters in length and occupied only a single line on the computer monitor. Each participant viewed 80 experimental sentences along with 72 filler sentences.

3. Results

In order to obtain a complete picture of the time-course of compound word processing, a number of dependent measures were examined. The primary measure was gaze duration, which is a sum of all fixation durations on the compound word prior to the eyes leaving

(F1 1 23 = 10 36, MSe = 2747, p = 004; F2 1 72 = 2 14, MSe = 4824, p = 148). There was an effect of beginning lexeme frequency, with compounds containing a highfrequency beginning lexeme receiving 36 ms shorter gaze durations than compounds with a low-frequency beginning lexeme (F1 1 23 = 13 67, MSe = 4388, p = 001; F2 1 72 = 5 34, MSe = 4824, p = 024). There was also a main effect of ending lexeme frequency in the analysis by participants. Compounds containing a high-frequency ending lexeme received 16 ms shorter gaze durations on average compared to compounds containing a low-frequency ending lexeme (F1 1 23 = 7 83, MSe = 1448, p = 01; F2 < 1). None of the interactions reached significance in this measure (all p’s > 15).

3.2. First fixations

First fixations landed 3.79 characters on average into the compound words. There were no significant main effects or interactions on first fixation position (all p’s >1).There was an effect of beginning lexeme frequency on first fixation duration, with compounds containing a high-frequency beginning lexeme receiving 8 ms shorter first fixation durations (F1 1 23 = 4 89, MSe = 720, p = 037; F2 1 72 = 1 77, MSe = 702, p = 188). There was also a marginal effect of ending lexeme frequency (F1 1 23 = 4 14, MSe = 1294, p = 054; F2 1 72 = 2 39, MSe = 701, p = 126) and a significant interaction between beginning and ending lexeme frequencies (F1 1 23 = 5 01, MSe = 1653, p = 035; F2 1 72 = 4 52, MSe = 702, p = 037). Follow-up comparisons demonstrated that the nature of the interaction was that compounds with two low-frequency lexemes received significantly longer first fixations (340 ms) compared with compounds in the other three conditions (321, 318, 316; all p’s<02 except comparing HH to LL for items where t2 40 = −1 97, p = 0 056), which did not differ significantly from each other (all t’s < 1).

3.3. Go-past duration

For go-past duration there was again a significant main effect of transparency, with opaque compounds taking 40 ms longer on average (F1 1 23 = 8 20, MSe = 9461, p = 009; F2 1 72 = 3 21, MSe = 9079, p = 078). In addition, there was a significant effect of ending lexeme frequency in the participants analysis, with compounds with highfrequency second lexemes producing 20 ms shorter go-past durations (F1 1 23 = 5 56, MSe = 3612, p = 027; F2 < 1). Transparency also significantly interacted with beginning lexeme frequency by participants (F1 1 23 = 6 69, MSe = 5066, p = 016; F2 1 72 = 1 81, MSe = 9079, p = 182). Specifically, while there was a significant 46 ms effect of beginning lexeme frequency for transparent compounds (t1 23 = −2 67, p = 0 014; t2 38 = −1 58, p = 122), there was no significant effect for opaque compounds (t’s < 1). These effects in go-past durations were not due to a difference in the percentage of