THE INFLUENCE OF TASK DEMANDS ON FAMILIARITY EFFECTS IN VISUAL WORD RECOGNITION: A COHORT MODEL PERSPECTIVE DISSERTATION

THE INFLUENCE OF TASK DEMANDS ON FAMILIARITY EFFECTS IN VISUAL WORD RECOGNITION: A COHORT MODEL PERSPECTIVE DISSERTATION Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University By Scott Steven Jankowski, M.A. * * * * * The Ohio State University 2006 Dissertation Committee: Approved by: Dr. Neal Johnson, adviser Dr. Mark A. Pitt Dr. John Opfer Adviser Psychology Graduate Program

ABSTRACT Familiar words are recognized more easily than are unfamiliar words. Readers become more familiar with words occurring frequently in the language, as well as with the letter-patterns that are common to many words. One measure of letter-pattern familiarity is the number of letter-positions within a word that can be changed to form another word. Two experiments examined the effects of changing task demands on how easily readers recognize words of varying levels of familiarity. In Experiment 1, readers made word/nonword decisions on words varying in their frequency of occurrence in written English texts and their number of letter-positions that yield more than one word. The response latency advantage for familiar words was greater when readers pronounced the words as opposed to a button-press response. This effect was also greater when wordlike nonwords were used in the task as opposed to unwordlike consonant strings. Additionally, words with many positions yielding multiple words delayed responses, but only in the context of wordlike nonwords. While consonant strings would be distinguishable from words almost immediately due to their highly irregular letter-patterns, wordlike nonwords would force readers to resolve each letter-position until either one candidate remained (in the case of a word) or all were eliminated (in the case of a nonword). Furthermore, readers responded to unwordlike nonwords more quickly in the context of words with many positions that could be changed to form other words, again indicating the readers sensitivity to this dimension of familiarity. ii

In Experiment 2, the subsequent recognition memory of the readers was tested. Pronouncing words during the lexical-decision task led to more reliable memory for the words, as did the inclusion of regular nonwords during the lexical-decision task. The latter observation reinforces the notion that wordlike nonwords force the readers to completely resolve each letter-position until a point of uniqueness is reached. These findings support a view of visual word recognition in which readers must resolve all of the letter positions in a word only under certain circumstances, depending on the demands of the particular reading task. iii

Dedicated to my wife and son iv

ACKNOWLEDGMENTS I would like to thank Neal Johnson for all of his patience, support, and encouragement. I would also like to thank John Opfer and Mark Pitt for their insight and suggestions. Thanks to Vincent Davis, Shellie Dubose, Dan Ford, Kim Moran, and Josh Vittie for their hard work during data collection. Finally, I would like to thank the members of the Johnson lab: Terri Childers, Virginia Gonsman, Leslee Martin, and Zach Schendel for making graduate school a bit more enjoyable. v

VITA August 7, 1973...Born Detroit, Michigan 1999 B.A. Psychology, Oakland University 1999- present.....graduate Teaching Assistant, The Ohio State University (OSU) 2002......M.A. Cognitive Psychology, OSU Major Field: Psychology FIELDS OF STUDY vi

TABLE OF CONTENTS Abstract ii Dedication. iv Acknowledgments... v Vita... vi List of Figures ix Chapters: 1. Introduction:.. 1 Page 1.1 Familiarity Effects: Word Frequency.. 2 1.2 Neighborhood Size as an Index of Familiarity. 7 1.3 A Cohort Model View. 13 1.4 Task Demands Affect Reading Strategies... 16 1.5 The Proposed Experiments 17 2. Experiment I 20 3.1 Participants.. 22 3.2 Materials and Apparatus. 22 3.3 Design.. 23 3.4 Procedure.... 23 3.5 Results.. 24 3.6 Discussion.. 37 vii

3. Experiment II.... 39 4.1 Participants....41 4.2 Materials and Apparatus....41 4.3 Design....41 4.4 Procedure...41 4.5 Results....42 4.6 Discussion.. 46 4. General Discussion.. 48 4.1 Summary....48 4.2 Ecological Validity of the Tasks.......50 4.3 Reconsidering Neighborhood Effects... 52 4.4 The Importance of the Number of Positions Yielding Neighbors...56 Bibliography..58 Appendix... 62 viii

LIST OF FIGURES Figure Page 2.1 Experiment 1: Interaction of number of positions yielding neighbors and nonword type for correct responses to words... 25 2.2 Experiment 1: Interaction of word frequency and nonword type for correct responses to words.....27 2.3 Experiment 1: Interaction of word frequency and response mode for correct responses to words.28 2.4 Experiment 1: Interaction of word frequency and number of positions at the regular nonword condition for correct responses to words..... 29 2.5 Experiment 1: Interaction of word frequency and number of positions at the irregular nonword condition for correct responses to words.... 30 2.6 Experiment 1: Interaction of word frequency and nonword type for correct responses to nonwords...32 2.7 Experiment 1: Interaction of nonword type and number of positions yielding neighbors for correct responses to nonwords... 33 2.8 Experiment 1: Mean number of errors for high and low frequency words.. 34 2.9 Experiment 1: Mean number of errors on word items in the regular and irregular nonword conditions.... 35 2.10 Experiment 1: Mean number of errors on word items in the spoken and buttonpress conditions..36 ix

Figure Page 3.1 Experiment 2: Mean d-prime scores for recognition task for participants in the regular and irregular nonword conditions during the lexical-decision task..43 3.2 Experiment 2: Mean d-prime scores for recognition task for participants in the spoken and button-press conditions during the lexical-decision task...44 3.3 Experiment 2: Mean beta scores for recognition task for participants in the spoken and button-press conditions of the lexical-decision task..... 45 x

CHAPTER 1 INTRODUCTION Models of visual word recognition have been shaped by studies of familiarity effects. A long-standing robust finding is that common words are read more easily than are rare words. Although this effect is widely considered to be due to the lexical entries of common words being more available during lexical selection, pre-lexical as well as post-lexical influences have been proposed. Despite these differences, common to all views of lexical access is the notion that familiarity with common words facilitates processing. However, the same notion should also extend to sub-lexical units, such as syllables or letter-patterns, with greater familiarity with the items facilitating processing at this level as well. The extent of a reader s familiarity with a word s letter-patterns may be indexed by neighborhood size, a measure commonly assumed to indicate the level of similarity a word shares with other words. The effects of large neighborhoods are commonly assumed to arise from competition between entries or the total amount of activation at the lexical level by words with large neighborhoods. However, if readers make use of sublexical units, then one possible contradiction that arises is that familiarity at these lower levels is associated with an increase in similarity to many other words. The 1

literature regarding neighborhood size effects has been inconsistent, especially when involving the lexical-decision task. Another neighborhood metric, the number of positions yielding neighbors, may be the cause of the confusing pattern of data. An account of familiarity effects will be offered in which letter-pattern familiarity facilitates processing at the pre-lexical level, word familiarity facilitates at the lexical level, and the number of positions yielding neighbors interferes with processing only under circumstances when the most careful scrutiny of the nonword items is required in order to distinguish them from real words. Furthermore, could the strategies and goals of the reader affect how well the words are later remembered? Memory for material that is more thoroughly processed should be more accurate. If the reader s task at hand requires only a passing glance, then the words may not be remembered well. On the other hand, if the reader must carefully consider each word, their memory for the words should be enhanced. Such an approach will provide behavioral clues to the nature of lexical access independent of the analysis of response times and error rates. Familiarity Effects: Word Frequency Although it is agreed upon that readers respond more quickly to words that appear more often in print, researchers have disagreed over the locus of this effect during lexical access. Solomon and Howes (1951) presented words tachistiscopically to participants and measured their identification thresholds for display duration. The words varied in the frequency they appeared in print according to the Thorndike-Lorge count (1944). They 2

found that common words yielded shorter identification thresholds. What remained to be determined was whether the effect was perceptually based, with common words being seen more easily, or due to differences in availability of the memory for high and low frequency words, with more common words being retrieved more easily. Addressing this issue, Duncan (1966) found that people tend to produce high frequency words when asked to generate an example that fits a few given constraints. Participants were either given some of the letters in the word or the class of object named by the word, and were then asked to name words that fit the constraints. This finding provides evidence that the word frequency effect is not entirely perceptually based because in this study the participants are presented with no stimuli, but instead must search their memories. In an early study on word frequency effects using the lexical-decision task, Rubenstein, Garfield, and Millikan (1970) reported faster response times to high frequency words. The authors proposed a model in which the stimulus is segmented, and consistent lexical entries are marked for comparison as more information arrives. The marking and comparison processes are assumed to proceed starting with high frequency words first, and to continue to lower frequency words after these are eliminated. This early model placed the effect late in lexical access, during the lexical selection stage. Landauer and Streeter (1973) questioned the validity of the word-frequency effect. They argued that many researchers assumed that rare and common words were structurally equivalent, and that word frequency effects were due to differences in frequency of usage. The authors demonstrated that rare and common words differ in the size of their orthographic neighborhoods and that they differ in their graphemic and 3

phonemic composition. They suggested that the apparent effect of word frequency might be due to these often-confounded dimensions. However, Gardner, Rothkopf, Laper, and Lafferty (1987) found word frequency effects when holding all the other dimensions constant and varying the readers exposure to the words. They constructed stimulus lists including words taken from the indices of engineering and nursing textbooks. Changing one or two letters from half of the items formed pronounceable and orthographically regular nonwords. They then tested a group of engineering students and professionals as well as a group of nursing students and professionals using the lexical-decision task. Each group responded more quickly to words that were common to their occupation. The difference was not found for the nonwords. The authors interpret the finding as evidence for a frequency based component of the word frequency effect, because even when all other word-level variables are held constant greater familiarity with a word improves performance. Balota and Chumbley (1984) questioned some assumptions concerning the locus of word-frequency effects. They argued that word frequency effects may not be due to an availability advantage enjoyed by common words during lexical identification, but may rather be due to post identification processes. They presented their participants the same set of words in the context of different tasks: lexical decision, pronunciation, and category verification. They found that the magnitude of the word-frequency effect varied among the tasks, with lexical decision demonstrating the greatest effect and category verification task showing no significant effects. Because all of the tasks were assumed to depend on lexical identification, the authors argue that identification must be insensitive to word frequency, and that the large effect often observed in the lexical-decision task is 4

task-specific. Balota and Chumbley (1985) suggested another argument against the location of word frequency effects in lexical identification. They presented words of varying frequency to their participants. They used a delayed naming task in which the participants were instructed to pronounce the words when prompted by a cue given after short delays of varying length. Since frequency effects were found after delays of 1400 ms, the authors argue that lexical access must have been completed and that word frequency must therefore influence production. Monsell, Doyle, and Haggard (1989) offered a criticism of these two studies. They pointed out that nine categories were used in the category verification task, which is substantially different than the two categories in a lexical-decision task. They also argued that the presentation of the category label at the beginning of each trial could have influenced the results because semantic priming attenuates frequency effects. They also suggested that category verification may not be dependent on unique identification, but could be influenced by prelexical semantic activation. If the reader could decide on the basis of the presence or absence of semantic activation consistent with the category label, no unique identification would be required. To address these issues, they performed an experiment comparing performance on a lexical-decision task to that on a modified semantic-categorization task. The same words were used for both tasks, and in the categorization task there was a choice between only two categories, with no category label presented before the presentation of the target. Significant effects of word frequency were observed for both tasks. Comparison of effects on the two tasks revealed no evidence of an interaction. When the specifics of the 5

categorization task more closely resembled those of the lexical-decision task, the frequency effect returned. The authors conclude that the categorization task is also sensitive to word frequency. As for the delayed-pronunciation task, Monsell, Doyle, and Haggard (1989) noted that the phonological characteristics of the words were not controlled in the Balota and Chumbley (1985) study. Furthermore, they also noted that responses were longer than a typical simple reaction time, suggesting that the readers were not completely prepared to give a response upon presentation of the cue, and may have needed to process some part of the stimulus again. The authors conducted a study that included a warning signal before the cue to permit adequate preparation of a response, the removal of the stimulus before the go signal, and a fixed interval to avoid uncertainty on part of the participants. Under these conditions, no effects of word frequency were found. The authors conclude that word frequency has little influence on post identification processes. Monsell, Doyle, and Haggard (1989) also addressed the comparison of word frequency effects in lexical-decision and naming tasks. They pointed out that pronunciation could depend on spelling-to-sound correspondences, and that word frequency should not influence the phonological assembly process. They compiled a list of words with both regular and irregular stress patterns. The items were intermixed to preclude the pronunciation based on assembly rules, thereby forcing lexical identification. The same participants also completed a lexical-decision task on the same set of words. The word-frequency effects were comparable between the two tasks. The authors argue that when naming tasks require lexical identification, the pattern of results resembles those found in lexical-decision tasks. 6

Jankowski (2002) took a similar approach in comparing the task demands in naming and lexical-decision tasks. In that study, participants completed a lexical-decision task in which the response mode of the participants was manipulated. Half of the participants made their responses by pronouncing the words and the other half made their responses by pressing a button. The requirement of pronunciation mandated that the participant choose a unique lexical entry prior to making a response. In the button-press condition, the participants did not have to settle on any particular word, but instead only needed to know whether or not the item in question was a word (i.e., any word). Because the spoken-response condition produced a larger word-frequency effect, the data supported the notion that high-frequency words enjoy an advantage during lexical selection. This study complements the Monsell, Doyle, and Haggard (1987) finding that making a naming task more demanding enhances the word-frequency effect. However, instead of comparing the magnitude of the effect across tasks, the demands of the lexicaldecision task were altered slightly in one condition in order to make the decision less dependent on access to a unique entry, which produced a relatively weaker effect. Neighborhood Size as an Index of Familiarity Neighborhood size was first used as a measure of a word s similarity to other words, with the expectation that a large number of words similar to the target would interfere with responses. Landauer and Streeter (1973) defined the orthographic neighborhood as the number of words that could be created by changing one position within a word, while leaving the others constant. Coltheart, Davelaar, Jonasson, and 7

Besner (1977) used neighborhood size to demonstrate shortcomings of serial search models of lexical access (Forster, 1976; Rubenstein, Lewis, & Rubenstein, 1971) in which the reader segments the incoming word, and compares the results to lexical entries in descending order of word frequency. For example, Forster (1976) proposed a model in which the lexical entries are searched serially until a match is found; much like a reader looks up an entry in a dictionary. The searches were assumed to run not in alphabetical order, but in order of relative word frequency in order to account for the word-frequency effect. The search would pause for verification when a sufficiently similar candidate was located, and would continue exhaustively until either a match was made (in the case of a word) or until no match was found (in the case of a nonword). The presence of similar many similarly spelled items would slow the responses due to a greater number of pauses to check similar items. Coltheart and colleagues (1977) demonstrated a shortcoming in these early word recognition models by demonstrating a neighborhood-size effect. The originally reported finding was a null effect for words, and an interference effect for nonwords. Because search models centered on a serial search of the lexicon in word-frequency order, large neighborhoods should interfere with both word and nonword responses. Coltheart and colleagues interpreted their findings as consistent with a logogen-style model (Morton, 1970) in which individual entries have evidence collectors (logogens) that receive activation in parallel. To account for the word-frequency effect, the threshold for access is assumed to be inversely related to the word s relative frequency. Because nonwords would have no evidence collectors, a deadline for making the nonword decision was 8

proposed. The interfering effect on nonword decisions was attributed to the reader s sensitivity to overall lexical activation. Nonwords with many neighbors would produce a high level of activation, as would the word stimuli, thus producing the neighborhood interference for nonword items. In this way, neighborhood effects were interpreted as interference due to similarity. The original report of neighborhood interference fit well with another class of models that shortly followed (i.e., the connectionist models). According to these models, upon visual presentation a word is assumed to provide activation to all similarly spelled items through activation of the lower levels of features and letters. Words with large neighborhoods would therefore activate many words. The best candidate is selected through means of lateral inhibition, with entries receiving more activation from the lower levels gradually inhibiting similar entries. The interactive-activation model (McClelland and Rumelhart, 1981) predicted interference from large neighborhoods due to the added time necessary to work through the interference. Later findings of facilitation by large-neighborhood items rather than interference (Andrews, 1987) conflicted with the explanation of interference by lateral inhibition. Words with large neighborhoods should activate a large number of similar candidates that will have to be deactivated. Andrews attempted to reconcile the data with the interactiveactivation model by proposing that neighborhood facilitation could be accounted for by feedback to lower levels from the lexical level. Pollatsek, Perea, and Binder (1999) later simulated a weak facilitation effect using a version of the interactive-activation model by slightly changing the word-letter excitation parameter. 9

To account for these new findings, new versions of the search model (Paap & Johansen, 1994), the logogen model (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001), and the interaction-activation model (Grainger & Jacobs, 1996) were proposed, each including one common additional feature. In these new models, as with previous versions, lexical access could be achieved when a single candidate entry was selected. What is new to these models is that they assumed that readers are also sensitive to the amount of overall lexical activity. Items will be judged to be words if the overall level of activation exceeds a predetermined threshold. This feature allows the models to account for the reported neighborhood facilitation for low-frequency words. Low-frequency words from large neighborhoods would produce sufficient activation to meet the criterion allowing a fast response to words. Andrews (1997) reviewed the scores of subsequent studies, and she noted that while neighborhood size always facilitated responses in a naming task, in lexical-decision tasks the effects ranged from that of facilitation through null to interference. Johnson & Pugh (1994) reported several findings that began to shed light on the confusing pattern of data produced by the lexical-decision task. They observed that large neighborhoods should correspond to a reader s greater familiarity with the word s letterpattern. In Experiment 3, the participants made lexical decisions on items that varied in terms of their neighborhood size as well as the number of positions yielding the neighbors. These two variables were positively correlated, as they are in the English language (and in all of the prior work done on neighborhood effects). When the participants had to decide between words and orthographically regular nonwords, more neighbors and positions produced slower responses. When the nonwords were 10

orthographically irregular, the words (from the same list) with more neighbors and positions yielding them produced faster responses. Further study was necessary to separate the unique effects of these two variables. In their fifth experiment, Johnson and Pugh (1994) controlled neighborhood size while manipulating the number of positions yielding neighbors. They used a lexicaldecision task with regular nonwords for foils. They found that words that had more letter positions yielding neighbors produced relatively slower responses. In this task, the words and nonwords were matched on both neighborhood size and number of positions. The nonwords were pronounceable and orthographically regular, so that they could potentially be English words, but they just happen not to be. Therefore, words would be indistinguishable from nonwords early in processing. They argued that since the participants could not use any of the early arriving information to make a lexical decision, each position had to be resolved to distinguish words from nonwords. Johnson and Pugh (1994) proposed a framework to account for these findings based on the COHORT model (Marslen-Wilson & Welsh, 1978) in which the speech signal is segmented, and the first arriving segment establishes an initial cohort of possible entries. Subsequently arriving segments deactivate incompatible cohort members until only the target remains. Although visual language is not spread out over time, as is speech, the encoding of each sublexical segment in a word is unlikely to occur simultaneously in printed language, even though the entire signal may arrive at the eye all at once. With more positions within a word yielding possible cohort members, the longer it should take to eliminate these possibilities, or to resolve the cohort. The reader is not expected to engage in this process under all circumstances, however. 11

Johnson and colleagues, (Johnson, et. al., 2005) further explored this notion by comparing performance across different tasks. In Experiment 4, participants completed a naming task in which they simply read the displayed items out loud. Neighborhood size and number of positions were independently varied. Large neighborhoods led to faster responses, and more positions yielding neighbors produced a null result. The naming task allows readers to assemble responses through phonological conversion. Therefore, cohort resolution was not required, and the positions effect was not observed. In Experiment 5, the same word lists were used in a lexical-decision task. Because assembling a response would lead to many incorrect responses, cohort resolution was necessary prior to initiating a response. While large neighborhoods again produced faster responses, more positions yielding the neighbors slowed responses. Jankowski (2002) investigated the same issue, but instead of comparing performance across tasks, he manipulated neighborhood size and the number of positions yielding neighbors, as well as the orthographic regularity of the nonwords within the context of a lexical-decision task. The purpose of the experiment was to observe the positions effect while changing the readers need for cohort resolution within the context of a lexical-decision task. The positions effect is only expected when cohort resolution is necessary. When regular nonwords were used, the Johnson & Pugh (1994) positions effect was replicated (i.e., more positions yielding neighbors resulted in slower response times). However, when irregular nonwords were used, the number of positions had no effect on response time. According to the cohort framework, word displays would produce lexical activity, while the nonwords would produce little activity, if any at all. The participants in the irregular nonword condition could take advantage of the early 12

arriving information regarding lexical activity that would permit a reliable lexical decision without the reader having to resolve the cohort. The manipulation of nonword type did not interact with the neighborhood-size effect, however. The stability of the effect suggests that neighborhood size influences processes early on in lexical access. The findings parallel those of the Johnson and colleagues (2005) experiments: neighborhood size facilitates responses regardless of the need for cohort resolution due to enhanced letter-pattern familiarity, and interference from the positions yielding neighbors depends on the need for cohort resolution, determined by the nature of the nonword displays. A Cohort Model View The concept of a cohort originated in the speech recognition literature (Marslen- Wilson & Welsh, 1978). According to their COHORT model, the listener segments the speech signal as it is received over time. The first arriving segment activates a subset of all lexical entries with which it is consistent, thus establishing the item s initial cohort. Subsequently encoded segments deactivate all cohort members with which they are inconsistent. Lexical access is therefore achieved when the cohort is resolved by narrowing down the original cohort to one member, when the point of uniqueness is reached. Due to the static nature of the signal, a cohort model in visual language may seem like a counterintuitive idea. Unlike the case of spoken language, in which the signal from a particular word is necessarily spread out over time, in written language the printed word 13

sits on the page. However, this apparent problem holds only if it is assumed that all of the information from the display is encoded simultaneously. Even if it were the case that all of the light reflecting from a printed word were to arrive at the retina at exactly the same time, there is no reason to believe that the user would encode all of the relevant information carried in the light at exactly the same time. If instead the signal is parsed and the segments are encoded gradually, the initial establishment and possible subsequent resolution of a cohort becomes possible. The cohort model of visual-word recognition (Johnson, 1992; Johnson & Pugh, 1994; Johnson, Childers, Jankowski, Miller, Gonsman, & Seifert, 2005) applies the cohort concept to the problem of reading. The model assumes that lexical access proceeds in a series of stages. Initially, the retina recodes the physical information in the light into a biological format. As segments of information become available, they are then recoded into the psychological formats of semantic, phonological, orthographic, and graphic information, in that order. These pieces of information, in turn, activate all lexical entries with which they are consistent, resulting in the item s cohort. Although many entries will be activated initially, those receiving activation from multiple sources will stand out. As more segments are encoded, lexical entries consistent with the new information would receive even more activation, while the activation would fade for those that are inconsistent with the new information. The model assumes that the reader will act as soon as sufficient information becomes available. If the reader s task is to simply pronounce a string of words, then the lexical entry that stands out early on could be selected with little chance of error. Because the reader correctly assumes that all of the to be encountered items are real words, the 14

entry with the highest activation level should most often correspond to the target. Words with more familiar letter-patterns will be processed more quickly, contributing to neighborhood size and word-frequency facilitation. However, a task that is often used in the reading literature, the lexical-decision task, introduces complications that may require the reader to resolve the cohort. For example, if the reader s task is to respond by pressing one button for words and another for the nonwords, then discriminating evidence will arrive earlier or later, depending on the composition of the nonwords. In the case when irregular nonwords, such as consonant strings, are used as foils, the reader could use the presence of an initial flush of lexical activation (in the case of the words) or the lack thereof (in the case of the nonwords) to decide without having to wait for one entry to stand out. However, when word-like (pronounceable and orthographically regular) nonwords are used, the reader will have to resolve the cohort in order to make the word/nonword decision. The level of initial lexical activity will not provide discriminating information, as it will be about the same for both words and nonword items. Furthermore, even nonwords will cause some particular lexical entry to stand out from the lexical noise, so simply choosing the most active member will result in many mistakes. Therefore, in this case, the reader will have to resolve the cohort by waiting until the encoding of further segments deactivates all of the other cohort members until either a single candidate remains (as in the case of a word), or all are eliminated (as in the case of a nonword). 15

Task Demands Affect Reading Strategies The previously discussed studies regarding the positions effect have indicated that readers may be able to adjust their reading strategies according to the demands of the task. Several other findings support this notion. Meyer and Schandeveldt (1971) reported that responses to a pair of words in a lexical-decision task were facilitated if the words were semantically related. Shulman and Davison (1977) later noted that they used orthographically regular nonwords in their task. They observed that only semantic information would differentiate words from nonwords in the task, and conducted another experiment using irregular nonwords. Under these conditions, the advantage for semantically related pairs was diminished. They argued that the participants could selectively monitor different streams of information about the items when making their decisions. Monsell, Patterson, Graham, Hughes, and Milroy (1997) sought to determine if readers could make use of different strategies within a naming task. Exception words, such as those with uncommon stress patterns, were chosen for the words. Some of the subjects were also exposed to nonwords intermixed with the words. The exception words must be pronounced by the activation of a lexical entry. Using an assembled-phonology process to pronounce these words would produce errors. However, the lexical route will not help the reader pronounce the nonwords, because nonwords do not have entries present in the lexicon. When the readers pronounced the words in pure blocks with no nonwords, they were faster and more accurate in their pronunciation than they were when nonwords were present in the list. This suggests that not only are both lexical and 16

sublexical routes present, but again that readers seem to make selective use of them depending on the particular demands of the task. The following experiments will manipulate task demands in a similar fashion in order to change the point at which the reader obtains information sufficient to make a reliable lexical decision. The Proposed Experiments Neighborhood size has typically been considered a measure of lexical similarity (Landauer & Streeter, 1973), meaning that words with large neighborhoods have many words that share similar spelling patterns. An influential connectionist model (McClelland & Rumelhart, 1981) predicts that a large number of similar candidates in the lexicon will interfere with the identification of a word. In their model, lexical entries receive activation from lower levels and laterally inhibit each other. The process of lateral inhibition will take longer if the target word has a large neighborhood. Furthermore, if lexical access proceeds in this fashion, then neighborhood-size effects should occur relatively late in the process, and should be sensitive to the particular task s demands for narrowing the field of candidates to a unique entry. However, another possibility is that words with large neighborhoods necessarily share more common spelling patterns. These commonly seen letter combinations could facilitate the encoding of the visual display very early on in lexical access (i.e., during perceptual encoding), regardless of the reader s task. The previously discussed findings support the notion that neighborhood size provides an index of letter-pattern familiarity. The facilitative effect of a large 17

neighborhood seems to be present regardless of the task or the particular task demands, as any reading task requires the initial encoding of letters. However, a large number of positions yielding neighbors can interfere with lexical access, but only when cohort resolution is required, as in a lexical-decision task which employs sufficiently word-like nonwords. Other tasks, such as a naming task, allow the reader to make use of other strategies such as using grapheme-to-phoneme correspondence rules to generate responses, or to select the most active lexical entry from the background of lexical activation. Under such circumstances, when only valid words are expected and encountered, the reader would not have to completely resolve the cohort. This notion may account for some of the contradictions apparent in the early studies on neighborhood-size effects. The null or inhibitory effects demonstrated in prior lexical-decision experiments may have been due to neighborhood size being confounded with the number of positions, with a result of the two effects having canceled each other out. Additionally, words that appear more frequently in print demonstrate a response time advantage. The magnitude of the effect varies with the nature of the task (Balota & Chumbley, 1984,1985; Monsell, et al., 1989). The cohort framework assumes that high frequency words will have more familiar letter-patterns due to the reader s increased familiarity with the pattern due to increased exposure. The facilitation due to ease of encoding relatively familiar letter patterns should be observed regardless of the particular reading task. Another source of this advantage is assumed to arise from a greater ease in picking frequently encountered items from the lexicon during the selection of a unique lexical entry. What is not yet clear is how the two late processes of cohort resolution and lexical selection interact. Prior studies examining the effect of positions have controlled 18

word frequency in order to avoid a potential confound. The reader should shift strategies depending on the demands of the task, and that the response times will shift accordingly. The cohort model predicts that if the reader must select a unique lexical entry, then the word-frequency effect should be relatively strong, as the high-frequency words will gain an advantage during both perceptual processing and lexical selection. However, instead of choosing the most active entry, if the reader will have to resolve the cohort, the process of cohort resolution could attenuate the word-frequency effect. If the cohort is narrowed to one remaining candidate, the target, high-frequency words may not gain an advantage under these conditions, as there would be only one entry to select, regardless of its familiarity. Furthermore, the type and amount of processing necessary to complete a lexical decision should not only affect the time necessary to make the response, and the accuracy of the response, but it should also affect the reader s subsequent memory for the words. If the readers vary their lexical access strategy based on the task demands, some of these strategies should lead to better memory performance than do others. The second experiment will provide converging evidence that the requirement of cohort resolution changes the reader s lexical access strategy. By changing the type of nonwords present during the lexical-decision phase, the resulting differential need for cohort resolution should produce predictable differences in memory performance. Readers in the irregularnonword condition may use incomplete information to make a lexical decision and should therefore demonstrate relatively poor memory performance compared to readers in the regular-nonword condition who must resolve the cohort in order to make a lexical decision. 19

CHAPTER 2 EXPERIMENT ONE The cohort model assumes that word frequency influences two stages of processing. High frequency of usage should increase the reader s familiarity with the word s letter-patterns. The encoding of sublexical units should be facilitated for highfrequency words. High-frequency words should also have an advantage at the lexical level, being more available for selection. The word-frequency effect should be weakened if the reader does not have to select a unique lexical entry. If readers give their responses by button-press, then they need not necessarily settle on a unique lexical entry prior to their decision. Participants who must pronounce the displayed words will be forced to choose a unique lexical entry, unlike those in the button-press condition who can initiate a response after any discriminating information becomes available. The manipulation of response mode (keyboard versus spoken responses) should change the reader s need for a unique lexical selection. This manipulation should attenuate the word-frequency effect, but not eliminate it, as readers in both conditions should respond more quickly to common words due to increased letter-pattern familiarity. The cohort model assumes that the facilitative effect of large-neighborhood words is also due to their relative ease of encoding due to enhanced familiarity with their common letter-patterns. However, a related effect, the number of positions yielding 20

neighbors, can attenuate or mask this effect if left uncontrolled in circumstances where cohort resolution is required. The need for cohort resolution will be manipulated by changing the type of nonwords present in the task. Participants who encounter orthographically regular and pronounceable nonwords should have to engage in cohort resolution in order to complete the task. Those who encounter unpronounceable consonant strings will get the benefit of early arriving discriminating information to make reliable lexical decisions (i.e., there is an activated cohort, or there is no activated cohort). In many of the prior experiments examining word-frequency effects and neighborhood effects, the number of positions yielding neighbors was left uncontrolled and it was correlated with neighborhood size. For example, Coltheart and colleagues (1977) reported a null neighborhood size effect for word items. Because the number of positions was not controlled in the experiment, the possibility exists that the facilitation from large neighborhoods was offset by the interference from many positions yielding neighbors. Furthermore, Andrews (1989) reported that the facilitative neighborhood-size effect interacted with word frequency, in that lower frequency words produced a more pronounced neighborhood advantage. This finding has become a cornerstone of models of lexical access. However, as the number of positions was left uncontrolled, it presents a potential confound. The apparent interaction of word frequency and neighborhood size may be influenced by the number of positions. In the current experiment, word frequency will be crossed with the number of positions yielding neighbors allowing the examination of whether the positions effect depends upon word frequency. Interactions observed between frequency and positions indicate the presence of a possible confound in studies in which the number of positions is left uncontrolled. 21

Method Participants. Two hundred fifty-six psychology students participated in the study. All spoke English as their native language. They received course credit for their participation. Apparatus. The items were presented to the participants on a Televideo 920C terminal controlled by a computer. The keyboard responses were collected using two keys on the terminal, and the spoken response times were collected using a microphone connected to a voice key. Materials. In order to manipulate the readers need for cohort resolution, two lists of nonwords were created. In one list, the nonwords were unpronounceable and orthographically irregular consonant arrays. An example is ptrw. In the other list, pronounceable and orthographically regular nonwords were used. These items were matched to the words in terms of both neighborhood size and the number of positions yielding neighbors, an example of which is nipe. One list of words was selected to cross word frequency and the number of positions yielding neighbors. Half of the words had high word-frequencies, averaging 106 occurrences per million according to the Kucera and Francis (1967) count, and the other half had low frequencies averaging about 5.5 per million. Half of the words chosen had two positions that yield neighbors, and the other half had four. The number of neighbors was balanced at an average of nine neighbors. The crossing of the two variables yielded four word sublists. All of the items used in the study were four letters long. The items appear in Appendix A. 22

Design. There are two between-subject variables in the study. The first is response mode. Half of the participants made their naming responses into a microphone, and half responded yes and no by button-press. The other between-subjects variable is the type of nonwords used in the stimulus lists. Half of the participants were exposed to the orthographically regular and pronounceable nonwords, while the other half were exposed to the orthographically irregular and unpronounceable nonwords. The within-subjects independent variables were manipulated using the four word sublists. Each sublist of words was combined with an equal number of nonwords to create the blocked lists. Each subject responded to all four lists. The order of the lists was counterbalanced, and the order of items within each list was randomized. Response latencies and error rates were recorded. Procedure. Single items were presented one at a time on the terminal. The items appeared white against a black background. An X marked the fixation point for 500 ms, after which it was replaced by a dot for an additional 500 ms. The test item immediately followed, and remained visible until the participant responded. The fixation point indicated the location of the second letter. Participants in the keyboard response condition were instructed to press a key labeled Y in response to words, and to press a key marked N in response to nonwords. In the spoken response condition, the participants were instructed to pronounce the words, and to respond to the non-words by saying no. A microphone triggered the computer to stop timing and record the response latency. All participants were instructed to make their responses as quickly as possible while making few mistakes. On every trial, participants were provided immediate feedback regarding their 23

accuracy on each item via a message on the terminal. Results Analysis of response times of correct responses to words Both task variables produced significant main effects. Irregular consonant strings produced faster responses than did the regular nonwords, F(1,240)=137.08, p <.001. This finding is expected as the decision is much easier with consonant strings used as foils. Also, button-press responses produced faster responses than did spoken responses, F(1,240)=130.78, p<.001. This is a typical finding as the spoken response not only requires a more physically complicated response, but also requires the reader to reach a unique lexical entry before the initiation of the response. These two variables did not produce a two-way interaction. The main effect of the number of positions yielding neighbors did not reach significance, F(1,240)=1.85, p >.05. However, as expected, the positions effect interacted with the type of nonword present in the list, F(1,240)=7.89, p <.01, with words yielding neighbors at a greater number of positions producing longer responses only when regular nonwords were used as foils. Figure 2.1 plots the interaction of nonword type and the number of positions yielding neighbors. Words with high word frequencies produced faster responses than did those with low frequencies, F(1,240)=386.03, p <.001. Word frequency also interacted with nonword type, with regular nonwords producing a more pronounced word-frequency effect, F(1,240)=87.67, p <.001. This was expected because the requirement of cohort 24

850 800 Response time (ms) 750 700 Four Positions Two Positions 650 600 Regular Irregular Nonword Type Figure 2.1: Interaction of number of positions yielding neighbors and nonword type for correct responses to words 25

resolution mandates the selection of a unique lexical entry. Figure 2.2 shows the interaction of word frequency and nonword type. Word frequency interacted with the reader s response mode as well, with spoken responses producing a stronger word frequency effect, F(1,240)=10.67, p <.001. When the readers are required to pronounce the target words, they must select a unique lexical entry prior to the initiation of their vocal response. High frequency words gain an advantage under these circumstances. Figure 2.3 shows the interaction of word frequency and response mode. The three-way interaction of word frequency, number of positions, and nonword type also reached significance, F(1,240)=5.23, p <.05. In the regular nonword condition, a greater number of positions led to slower response times. However, in the irregular nonword condition, a greater number of positions led to faster response times for low frequency words. Figure 2.4 shows the interaction of word frequency and the number of positions at the regular nonword condition. Figure 2.5 shows the interaction of word frequency and number of positions at the irregular nonword condition. Furthermore, the three-way interaction of word frequency, number of positions, and response mode also reached significance, F(1,240)=5.80, p <.05. In the button-press condition, a greater number of positions led to slower responses. In the spoken response condition, a greater number of positions led to faster response times for low frequency words. Analysis of correct responses to nonwords The main effect of word frequency was significant, F(1,240)=48.66, p <.001, with participants making decisions about nonwords more quickly when high-frequency words were present in the lists. The frequency effect interacted with the type of nonword, 26