LANGUAGE AND COGNITIVE PROCESSES, 2003, 18 (3), 335–362 Morphological structure in the lexical representation of prefixed words: Evidence from speech errors Alissa Melinger Max Planck Institute for Psycholinguistics, Nijmegen, The Netherlands The present study aims to determine whether semantic relatedness plays a role in the production of speech errors involving derivational morphemes. A word order competition technique was used to induce morpheme and syllable exchange errors. Semantic relatedness was manipulated by contrasting error rates for prefixed words derived from free stems to those derived from bound roots. Significantly more morpheme errors were elicited in the two prefixed conditions compared with the control condition, replicating prior findings in French. Crucially, the two prefixed conditions elicited equal numbers of morpheme errors and there was no correlation between semantic relatedness and error rates. Taken together, the results strongly support a model of speech production in which derivational morphemes are represented at the form level and are not influenced by the degree of semantic relatedness within a morphological family. INTRODUCTION One goal of morphological theory is to characterise the knowledge that speakers possess about the morphological structure and relatedness of words in their language. The ability to understand and create new words composed of familiar parts, clearly an important component of linguistic Requests for reprints should be addressed to Alissa Melinger, Max Planck Institute for Psycholinguistics, Postbus 310, 6500 AH, Nijmegen, The Netherlands, Email: [email protected]. This research was supported (in part) by research grant 1 RO1 MH60133-01 from the National Institute of Mental Health, National Institutes of Health awarded to Gail Mauner and Jean-Pierre Koenig. I would like to thank Jean-Pierre Koenig, Karin Michelson, Gail Mauner, Paul Luce, Wendy Baldwin, Eve Ng, Anne Vestergaard, Breton Bienvenue, Asifa Majid, Karin Humphreys and two anonymous reviewers for their helpful comments and support. c 2003 Psychology Press Ltd http://www.tandf.co.uk/journals/pp/01690965.html DOI: 10.1080/01690960244000072 336 MELINGER competence, draws on the knowledge of a language’s morphology. While morphological theories aim to characterise what speakers know about the relationships between words, models of speech processing continue to debate whether morphological information plays any role in producing or recognising words. One debate within the processing literature is whether a morphological level of representation is needed; many researchers argue that form and meaning levels, which are independently required, can accommodate all the apparent morphological effects without postulating an additional explicit morphological level of representation. Most morphologically related words have both formal overlap, sharing some phonological form, and meaning overlap, sharing some meaning components. It is well known that words with only form similarity (Tanenhaus, Flanigan, & Seidenberg, 1980) or only meaning similarity (Neely, 1976; Swinney, Onifer, Prather, & Hirshkowitz, 1979) do produce facilitative effects independent of morphological relatedness. Therefore, it is possible that purported morphological effects emerge as the result of the combination of the independent form and meaning effects and not as the result of a morphological relationship, per se. Indeed, several researchers have argued that experimental effects which are attributed to morphological processes can also be explained by differences in degrees of semantic similarity (Plaut & Gonnerman, 2000) or phonological overlap (Rueckl, Mikolinski, Raveh, Miner, & Mars, 1997). While the debate swings back and forth in the speech recognition literature, within the speech production literature, arguments for an independent morphological level of processing have only recently been addressed (Pillon, 1998; Roelofs, 1996; Roelofs & Baayen, 2002; Zwitserlood, Bölte, & Dohmes, 2000). The aim of the current study is to provide additional support for an explicit level of morphological representation as well as to examine the role of semantic relatedness to morphological processing in speech production using an experimental elicitation of speech errors. Speech errors have been used for decades to form inferences about the nature of linguistic representations (Fromkin, 1971). The distribution of various types of errors was a primary source of evidence for the formation of speech production models (Dell, 1986; Garrett, 1980, 1982). For example, error data were central to the now widely accepted proposal for a separation between a level of semantic/syntactic representation and a level of form representation. Misordering errors such as in (1a) and (1b) occur at the stage of production when the phonological form of the word is being retrieved and assembled. These errors also provide external evidence for linguistic notions such as phonological features, segments, syllables, etc. Errors involving the mis-selection of a semantically related word, such as in (1c)–(1e), occur at an earlier level of processing when the appropriate MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 337 lemma is selected to express a particular concept. Errors of this type argue for a semantic organisation of the mental lexicon. (1) a. skip one stage Õ stip one skage b. Pity the new teacher Õ mity the due teacher c. Blond hair Õ blond eyes d. My dissertation is too short . . . long e. She’s a real [swKp] chick Segmental error respecting syllabic position Featural error Semantic error Semantic word substitutions Semantic blend between swinging and hip Speech errors can also involve the misordering or mis-selection of morphemic units, most frequently inflectional morphemes. Inflectional morphemes can be mis-selected, as in (2a), or they can be stranded when their stems are misordered, as in (2b). (2) a. he had to have it Õ he haved to have it b. a floor full of holes Õ a hole full of floors Derivational morpheme errors have also been reported, however they are much less frequent than inflectional errors (Cutler, 1980; Fromkin, 1973; Garrett, 1980; MacKay, 1980; Stemberger, 1985). Despite the prevalence of morphological errors in the literature, the interpretation of morphological slips, especially derivational errors, remains unclear. Derivational morpheme errors, such as those in (3), are difficult to interpret as necessarily involving morphological units. For example, in (3a) the error could be an example of a phonological word blend, perhaps between grouping and arrangement. Likewise, in (3b) the error could be the result of a phonological word blend between the target inquisitive and a competitor exquisite. Under this interpretation, the errors are no different from other word blends between two non-morphologically complex words, such as in (4). In (3c), it is possible that the speaker incorrectly applied a word formation rule when the existing word was temporarily unavailable, creating the error derival rather than retrieving derivation. While this error does inform us that words can be productively produced on-line, it tells us nothing about the representation of derivation or of the role of morphemes as units in speech production. (3) a. b. c. Intended grouping inquisitive derivation Produced groupment exquisitive derival 338 (4) a. b. MELINGER Intended public/popular person/people Produced poplic perple One way to differentiate between a morphological explanation and a phonological word blend explanation for errors that appear to involve morphemes is to investigate the distribution of naturally occurring errors. For example, there is evidence to suggest that inflectional suffixes pattern differently in errors than other non-morphemic word endings, which supports a morphological interpretation of inflectional errors (Bybee & Slobin, 1982; MacKay, 1976; Stemberger & MacWhinney, 1986). Unfortunately, the same sorts of distributional and probabilistic evidence for derivational affixes is sparse. Another way to discriminate between the two possible explanations is to experimentally induce errors. For example, MacKay (1978) induced errors by eliciting derived forms of simple words from speakers. Speakers often produced an error which involved the over-generalisation of a morphological rule, such as in (5). However, critics have argued that this elicitation method is artificial and does not reflect the same processes involved in naturally occurring errors since these errors do not involve the retrieval and blending of existing words. As a result, MacKay’s findings may not bear on the lexical representations of existing complex words but potentially only on the production of new words via rule application. (5) collide Õ collidement, not collision The existence of naturally occurring derivational errors such as in (3) have been used as evidence that roots and derivational affixes are stored and accessed separately (Fromkin, 1973:236). However, the alternative interpretation which denies the active involvement of derivational morphemes in errors is still possible; strong conclusions cannot be drawn from the data that come from naturally occurring errors. To provide stronger support for a morphemic analysis of derivational errors, Pillon (1998) demonstrated that derivational units participate in errors more frequently than non-morphological units. Pillon conducted a Word Order Competition (WOC; Baars & Motley, 1976) study to elicit derivational morpheme errors. Originally, the WOC technique had been used to induce spoonerisms (e.g., darn bore becomes barn door; Baars & Motley, 1976), syllable exchanges (e.g., horrible miracle becomes horracle mirible; Baars, Matteson, & Cruickshank, 1985) and word exchanges between phrases (e.g., He fixed his trousers and dropped his watch. Italicised words are exchanged; Baars & MacKay, 1978). Pillon used the WOC task to elicit morphological and phonological errors to determine whether morphemic MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 339 units were more likely to participate in speech errors than non-morphemic units. Pillon (1998) presented word pairs consisting of a French noun þ adjective. Her presentation word pairs were either both suffixed, such as in (6), or both monomorphemic words with phonologically matched pseudosuffixes, as in (7). (6) a. a. b. (7) a. b. c. troupeau traı̂nard penseur hautain crochet mural cadeau bâtard auteur soudain chalet rival ‘flock straggler’ ‘thinker haughty’ ‘hook wall’ ‘gift hybrid’ ‘author unexpected’ ‘chalet rival’ These word pairs were presented to participants who were instructed to pronounce the words either in the order in which they appeared or in the reverse order. The ordering uncertainty gives rise to a high proportion of speech errors (Baars & Motley, 1976). In the case of morphological-type errors, ordering uncertainty should produce exchange errors such as in (8)– (9). The expected errors in (8) produce one real word, an intruding word, and one nonsence word; Pillon also used examples that produced two nonsense words, such as in (9). In each of these errors, the (pseudo)roots are exchanged and the (pseudo)affixes are stranded. Pillon hypothesised that ‘‘derived word pairs should give rise, more often than control ones, to such stranding errors, if lexical retrieval actually entails the selection of intermediate morphemelevel units between words and phonemes’’ (1998:472). Presented pair (8) a. troupeau traı̂nard ‘‘flock straggler’’1 b. cadeau bâtard ‘‘gift hybrid’’ (9) a. chaton froussard ‘‘kitten cowardly’’ b. flacon standard ‘‘bottle standard’’ Expected error traı̂neau *troupard ‘‘sleigh’’ bâteau *cadard ‘‘boat’’ *frousson *chatard *standon *flacard Pillon reported significantly more stranding errors for morphologically complex words than for monomorphemic control words. She concluded that ‘‘morpheme units like roots and derivational suffixes are handled at some stage of the real-time speech production process’’ (1998:487) and that derivational structure needs to be represented in the speech system. 1 Literally ‘‘slow coach’’. 340 MELINGER Pillon’s study, together with the naturally occurring derivational errors, provides compelling evidence for the inclusion of derivational morphological structure in the lexicon. However, one alternative explanation for Pillon’s results is that semantic relatedness mediated the production of morphological errors. It is well known that word substitutions are more common when target and error are semantically and phonologically similar compared with just being phonologically similar (e.g., I thought Westerns were where people ride horses instead of cars Õ . . . instead of cows; Fromkin, 1973; Dell & Reich, 1977, 1980). Given the very nature of morphemes, words which share a stem or an affix will have more semantic similarity than two words sharing a syllable. The presentation and intruding words in Pillon’s suffixed condition had more semantic overlap than in the pseudosuffixed condition. Therefore, it is possible that the elicited ‘morpheme’ errors were actually phonological blends whose frequency of occurrence, or probability of occurring, was enhanced by the additional semantic relatedness between the presentation and intruding words. The possibility that Pillon’s exchange errors could have been influenced by semantic factors is supported by the fact that the biased pairs, those designed to produce errors consisting of one real word, such as in (8) above, produced significantly more errors than the unbiased pairs, designed to produce errors consisting of no real words, as in (9) above. Errors that produce real words are more common than errors which do not; thus, the difference in exchange error rates between her biased and unbiased pairs is not surprising (cf. Baars, Motley, & Mackay, 1975). However, given that a semantic explanation for the errors is strongest when at least one intruding word is competing for selection, this difference adds to the uncertainty about the mechanism and representation underlying these errors. Pillon maintains that semantic relationships between presentation and intruding words do not influence the observed morpheme exchange errors because many of her words were, in fact, not semantically related. However, her measure of semantic relatedness was intuitive, not objective. Intuitive judgements of semantic relatedness, especially from non-naive speakers, are problematic and often differ greatly from objective judgements obtained with ratings studies. In fact, pairs that Pillon intuitively classified as having ‘‘at best only a loose semantic and morphological relationship’’ were classified by several native French informants as transparently related (J.-P. Koenig, personal communication, May 18, 2002). Thus, the role of semantic relations in influencing morpheme exchange errors remains open. To determine whether morpheme errors occur independent of a semantic relationship, a Word Order Competition (WOC) study was MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 341 conducted comparing morpheme exchange errors produced for English prefixed words with different degrees of semantic relatedness. Specifically, errors were elicited to English prefixed words derived from free stems (e.g., reload ~ unload), which have a relatively high degree of semantic relatedness within a morphological family, and from words derived from bound roots (e.g., induce ~ reduce), which have a relatively low degree of semantic relatedness within a morphological family. The number of exchange errors produced for these two types of complex word, as well as the number of phonological and lexical errors, were compared with the number produced in a phonological control condition. Only semantic relatedness, as a function of the type of base the word is derived from, was systematically manipulated. If semantic relatedness plays no role in the production of morpheme exchanges, then both prefixed words derived from free stems and bound roots should produce more exchange errors than the control condition and they should not differ from each other. In contrast, if semantic similarity is a component which underlies the production of exchange errors, then only the free stem words should produce more exchange errors than the control condition. This result would argue against the retrieval of explicitly represented morphological constituents during speech production. EXPERIMENT In the WOC technique, pairs of words are presented to participants for a naming response. Exchange errors involving word components are elicited by randomly changing the response instructions such that participants are unaware when reading the words which order they will have to be named. The ordering uncertainty induces high error rates. The number of exchange errors in three conditions of presentation words are compared: free stem words, bound root words, and phonological control words. Because there is a strong lexical bias in error corpora, i.e., errors that result in actual words are more frequent than errors that produce nonsense strings, only materials that produce real word errors were included (Baars et al., 1975). Unfortunately, the constraint against nonsense errors restricts the possible construction of morphologically simple words with matching pseudoprefixes. Therefore, unlike Pillon’s monomorphemic control group which was phonetically matched to the complex group, as shown above in examples (6)–(7), the current control condition consists either of morphologically simple words (without pseudoprefixes), as in (10), or suffixed words, as in (11), with exchangeable syllables. The disadvantage of this design is that the items in the control set are not phonologically matched to particular complex word sets. The advantage, however, is that the entire set of stimuli can contribute to an 342 MELINGER effect since they all respect the strong lexical bias for errors. Crucially, for the suffixed simple words, the exchangeable syllables are not co-extensive with morphemes. (10) a. b. (11) a. b. Presented pair socket pallid rattle morbid scanty pedal certify forgery Expected error solid packet mortal rabid scandal petty surgery fortify To evaluate the influence of semantic similarity on experimentally induced morphological exchange errors, the number of morphological exchange errors produced by prefixed words derived from free stems was compared with the number produced by prefixed words derived from bound roots. As a baseline measure of syllable exchanges, the number of syllable-level errors produced in the control condition was also compared. Only errors involving the relevant subportions of words, namely morphemes and syllables, should be influenced by the manipulation of word type; all other errors should be produced at equal rates across conditions. METHOD Participants. Forty-two native English speaking undergraduates from the University at Buffalo, State University of New York received partial course credit for their participation. Materials. Thirty-three experimental word pairs were constructed for presentation in a Word Order Competition (WOC) study.2 Complex word pairs consisted of two prefixed words, derived either from free stems or bound roots. Control word pairs consisted of two non-prefixed words. The words in each presentation pair were semantically, morphologically, and phonologically unrelated to each other. An example set of presentation word pairs along with the expected intruding errors are presented in (12). To increase the rate of morpheme and syllable exchanges, all word pairs were designed to produce real word errors. All intruding word errors in the prefixed conditions were morphologically complex. (12) Free Bound Control Presentation pairs enfold react reject induce rattle morbid Possible error enact refold reduce inject rabid mortal 2 Originally, 13 item sets were constructed for each condition. However, two sets in the control condition were found to not meet the necessary criteria for inclusion in the study. Therefore, those item sets, as well as two randomly selected item sets from each of the other conditions, were excluded from all analyses. MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 343 In the two prefixed conditions, the word components targeted for the exchange error correspond to English morphemes, namely, prefixes and bases. A variety of prefixes were used. The free stem condition included seven different prefixes and nine different prefix combinations (only the pair dis - re was repeated). The bound root condition included 12 prefixes (14 different prefix allomorphs) and 11 prefix combinations. Between the two prefixed conditions, three common prefixes were used.3 For all words in the two complex conditions, the splicing point, i.e., the point at which the words can divide and exchange parts to produce real words, occurred at the morpheme boundary. The morpheme boundary coincided with a syllable boundary in 10 of the 22 word pairs. All of the control condition words conformed to the following criteria: (1) the two exchangeable word fragments which remain after the word has been divided do not constitute real words in English nor do they constitute roots or affixes that are semantically related to the meaning of the whole word; (2) the two word fragments correspond to the syllables of the presentation and intruding words, not just possible English syllables. The control condition consisted mostly of monomorphemic words; seven of the 22 words in this condition contained a derivational suffix and one contained an inflectional suffix. Furthermore, one pair included two complex words; the remaining 10 word pairs consisted either of two simple words or of one simple and one suffixed word. Importantly, when a control word was suffixed, the splicing point always occurred at the syllable boundary, not at the morpheme boundary; the exchangeable syllables were not coextensive with morphemes. Additionally, the exchange of morphemes between two words in the control condition did not result in a legitimate word (e.g., certify forgery Õ *certery *forgify) while an exchange of syllables did (e.g., certify forgery Õ surgery fortify). In an attempt to control for syllable frequency, the words in the control condition often consisted of high frequency word endings. Of the monomorphemic words in the control condition, 14 ended with (pseudo)suffixes and the remaining eight were composed of high frequency endings such as -ow, -le, or -el. In 6 of the 11 word pairs, the potential errors were phonologically but not orthographically matched (e.g., fashion setter Õ fatter session and certify forgery Õ surgery fortify). Evidence from prior studies using this 3 Most prefixes could not be used in both conditions because they are restricted in their distribution. Certain affixes, which are sometimes termed ‘‘level 2’’ affixes like un- and mis-, only occur with free stems while ‘‘level 1’’ affixes, like ad- and per-, only occur with bound roots. 344 MELINGER task suggests that conflicting orthography should not affect error rates as long as the phonological representations are consistent (Baars & Motley, 1976). Of the 66 words that comprised the 33 pairs across conditions, 54 were bisyllabic, and 12 were trisyllabic. Baars et al. (1985) showed that syllable exchanges are more likely to occur between words that have the same number of syllables and stress patterns. Each word within a pair had an equal number of syllables and the same stress pattern. Each word pair in all conditions met the following criteria: (1) exchanging morphemes/syllables results in the production of actual English words; and (2) no target or expected error was produced by exchanging only a single phoneme (i.e., the prefixes pre and re were never paired since a prefix exchange and a segment exchange would be indistinguishable). In addition to the 33 experimental word pairs presented, 72 distractor word pairs were interspersed with the experimental items. Distractors consisted of random word combinations as well as rhyming words (e.g., heel kneel), alliterative words (e.g., greedy green), semantically associated words (e.g., yellow white), and words standing in unusual syntagmatic relationships (e.g., man hairy). All items were distributed within a single list, resulting in a fully within-subjects design. Conditions were matched on two measures of word frequency: the raw (and log) frequency summed across a presentation word pair and the difference between the presentation pair’s summed frequency and the intruding (or error) word pair’s summed frequency. A one-way analysis of variance revealed no difference in word frequencies across conditions, F 5 1, but a marginal difference when comparing the difference in frequency between presentation and intruding word pairs, F(2, 30) ¼ 3.2, MSe ¼ 1102, p 5 .06. This marginal difference was due to higher intruder frequencies in the bound root (61.7) and control conditions (37.1) compared with the free stem condition (17.5). Because word frequency influences word-form retrieval times (Jescheniak & Levelt, 1994), intruding words were designed to have higher word frequencies than the presentation words to induce higher rates of exchange errors in all three conditions. Since intruders in the bound root and control conditions have higher frequencies, it is possible that the number of errors in these conditions will be enhanced. This possibility will be addressed in the discussion. The overall phonological similarity between the two words forming a presentation pair was balanced between conditions. The phonological similarity was measured grossly, in terms of shared segments, not shared features. A segment was counted as shared if it was common to both words of a presentation pair and held the same syllabic position in both words. MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 345 For example, the words insist conclude have one segment overlapping, namely the [n] in the coda of each initial syllable. In contrast, the words concede profess have no segmental overlap despite the fact that both words contain an [s]. Since the [s] in concede is in a syllable onset and the [s] in profess is in a syllable coda, they are not coded as overlapping. Given this coding system, the average number of overlapping segments within a condition was calculated. A one-way analysis of variance revealed no significant difference between conditions, F 5 1. It was impossible to fully match string length across conditions. There was a significant difference in the graphemic string length between conditions, F(2, 30) ¼ 9.43, MSe ¼ 3.3, p 5 .01. This difference was graded between the three conditions, the free stem condition being the longest with 7.7 graphemes per word, followed by the bound root condition with 6.65 graphemes per word, and the control condition with 6 graphemes per word. Additionally, while the two words in a pair always had an equal number of syllables and the same stress pattern, there was some variation in the number of syllables between item sets. The free stem condition had an average of 2.5 syllables per word, the bound root condition had an average of 2 syllables per word, and the control condition had an average of 2.1 syllables per word. This difference, carried by the length of the free stem words, was also significant, F(2, 30) ¼ 5.25, MSe ¼ 0.12, p 5 .05. Free stem and bound root words are used in this study as a method of comparing error rates for prefixed words with differing degrees of semantic relatedness. However, while bound root words generally do not have strong semantic consistency across their morphological families, they may still have more semantic overlap than the control words. For example, while the relatives receive and deceive are semantically unrelated, the relatives increase and decrease are highly related. Likewise, despite having a clearer semantic composition, free stem words can also have opaque relations to morphological relatives. For example, recharge and discharge are not as clearly related in meaning as reappear and disappear. Therefore, a semantic rating study was conducted to ensure that the conditions differed in their degree of semantic relatedness. Sixty-six word pairs were constructed for a semantic relatedness rating study. Each individual presentation word for the WOC study was paired with its predicted intruder word. For words from the prefixed conditions, the intruder word was composed of the shared stem or root and the prefix from the other word in the relevant presentation pair. For the control condition, the intruder was a rhyming word with the first syllable taken from the other word from the relevant presentation pair. Examples are given in (13). In addition to the 66 experimental word pairs, 180 distractor word pairs were presented. 346 (13) MELINGER Condition Free Presentation pairs Recharge Displace Bound Insist Conclude Control Miller Parcel Rating study pairs Recharge Discharge Displace Replace Insist Consist Include Conclude Miller Parlor Missile Parcel4 Rating 3.2 3.7 2.2 4.2 2.2 2.0 Average 3.45 3.2 2.1 Word pairs were centred on the computer monitor with a 7-point Likert scale at the bottom of the screen. Participants were asked to rate the degree to which the words were related in meaning.5 Lower ratings signify that participants did not identify a clear meaning relationship between the words while a high rating indicates that they did identify a clear meaning relationship. Semantic similarity ratings were calculated for each rating study pair and compared across conditions. Additionally, a second measure of semantic similarity was calculated by averaging together the two individual ratings to obtain a rating for each presentation pair. The rating study revealed that the conditions do differ in the degree to which presentation words and intruding words are semantically related. The free stem condition had the highest average rating, 4.15, followed by the bound root condition, 3.5. The control condition had the lowest average rating, 2.4. This difference was significant across the three conditions by subjects, Friedman: w2(2, N ¼ 10) ¼ 14.6, p 5 .01, and items, Kruskal–Wallis: w2(2, N ¼ 66) ¼ 32.59, p 5 .01. Paired comparisons reveal that the gradient difference was significant between all three conditions. The free stem words received higher relatedness ratings than the bound root words, Wilcoxon Sign: z ¼ 2.5, p 5 .05; Mann–Whitney U: z ¼ 2.5, p 5 .05, for subjects and items, respectively, and in turn the bound root words received higher relatedness ratings than the control words, Wilcoxon Sign: z ¼ 2.7, p 5 .01; Mann-Whitney U: z ¼ 3.7, p 5 .01, for subjects and items, respectively. Procedure. The procedure used was an adaptation of the Word Order Competition (WOC) technique of Baars and Motley (1976). Each trial began with the presentation of a word pair centred on the computer screen in capital letters. Each pair of words remained on the screen for 900 ms. After a 200 ms delay, a response cue was presented. There were two 4 In the mid-western and New York dialects spoken by the participants in this study, the word missile has only two syllables, ['mI.s@l], and rhymes with parcel ['par.s@l]. 5 In the instructions, participants were informed that words could be related in a number of different ways such as synonymy (SOFA–COUCH), autonymy (HOT–COLD), or associatively (TEACHER–STUDENT). They were instructed to decide for themselves whether two words were related, or similar, in meaning and to rate the degree of relatedness. They were also encouraged to use the entire range of the scale, not simply the two endpoints. MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 347 different response cues; REPEAT and REVERSE. When the response cue was REPEAT, participants were instructed to name the words in their presented order. When the response cue was REVERSE, participants were instructed to name them in the reverse order. All experimental targets were followed by the REVERSE response cue. Removing the word pairs from the monitor before the presentation of the response cue prevented participants from simply reading the words off the screen. The response cue remained on the screen for 1 second after onset of response. Following this interval, the next pair of presentation words was displayed. Participants sat before a computer monitor and a microphone. There was also a tape recorder in the room. Trials were recorded and responses were transcribed by two raters, one of whom was blind to the hypotheses of the study. The inter-rater reliability was 97%. Errors were then classified into types. RESULTS AND DISCUSSION Classification of errors The most common errors were skipped trials, incomplete responses, failure to reverse the word order, mispronunciations (various phonological errors), external intrusions, inflectional errors, and derivational exchange errors. All but the exchange errors should be unaffected by the current manipulation. Following the classification schema of Pillon (1998), errors were classified either as contextual or non-contextual. Contextual errors involve misorderings of segments present within the trial. Non-contextual errors involve the omission or insertion of segments not present in the trial. External intrusions, incomplete responses, and inflectional errors are classified as non-contextual errors. Contextual errors were further classified as within-word and between-word errors. Within-word errors involve a misordering of segments within a word while between-word errors involved a misordering of segments between words. For example, rattle morbid Õ battle morbid is an example of a between-word contextual error because features from the second word are anticipated in the first word. In contrast, misnamed unused Õ misnamed unsued is a within-word contextual error since segments from within the second word are misordered; the first word plays no role in the error. Finally, betweenword contextual errors were further divided into critical and (sub)segmental errors. Critical errors involve the portions of the words that are investigated in the current study, namely morphemes or syllables. (Sub)segmental errors involve between-word exchanges of phonological segments, features, etc. Thus, phonological exchanges involving single features or segments, such as in rattle morbid Õ battle morbid, are classified as (sub)segmental between-word contextual errors while an exchange of 348 MELINGER Figure 1. Hierarchy of error type classification. prefixes, such as in reject induce Õ reject reduce, is classified as a critical between-word contextual error. Critical errors not consisting of two full words were further classified as incomplete errors. The hierarchy of error types is depicted in Figure 1. Examples of the different error types with their classifications are presented in Table 1. Distribution of errors The manipulation of word type is predicted to only affect the production of critical between-word contextual errors. The distribution of non-critical errors, i.e., all errors that do not involve the critical components of the words under investigation, is predicted to be unaffected by the current manipulation. There are three types of critical errors: prefix/initial syllable spreading, where both words are produced with the same word onset, such as in (14), stem/final syllable spreading, where both words share the same word ending, such as in (15), and total exchanges, where both word onsets and offsets are swapped, such as in (16). All spreading errors can occur by perseveration or anticipation of the morpheme or syllable. In order for a speech error to be included in the analysis of critical errors, the production of the word pair must be complete. Partial errors and incomplete responses, even those that clearly include a morphological error (reject induce Õ reject re . . . or reject induce Õ reduce . . .), were not included in the analysis. Correct response (14) reject induce (15) reject induce (16) reject induce Error reject reduce reject inject reduce inject Error type Prefix spreading Stem spreading Total exchange 349 reject induce dispose propel reject induce rattle morbid remit advise misnamed unused dispose propel unordered discounted insist conclude perturb dissuade misprint prejudge reject induce reject induce reject induce Failure to reverse word order Failure to complete the pair Inflectional errors Derivational errors External intrusions Phonological errors Target utterance Error type induce reject dispose reduce battle morbid remit admise misnamed unsued dispose [próp@l] unordered discouraged interest conclude perturbed dissuade misprint prejudice reject reduce reject inject reduce inject Produced error TABLE 1 Examples of errors organised by error type Non-contextual Non-contextual Contextual, between, critical, incomplete contextual, between words, (sub)segmental contextual, between words, (sub)segmental contextual, within words Non-contextual Non-contextual Non-contextual Non-contextual Non-contextual contextual, between words, critical, complete contextual, between words, critical, complete contextual, between words, critical, complete Classification 350 MELINGER Each of the 42 participants contributed a maximum of 33 responses, creating 1386 opportunities for error, 462 per condition. Because the task was fast-paced, participants were occasionally unable to make a response. In total, 54 trials (less than 4% of all experimental trials) were skipped by participants, and 79 trials (5.7%) were incomplete (without an error). Most incomplete responses consisted of only one of the two response words. There were an additional 32 partial completions which included a morphological error, 23 from the free condition, 7 from the bound condition and 2 from the simple condition. In total, 12% of the trials were lost due to incomplete responses. In total, 43% of the responses had some type of error. The overall distribution of errors by condition is presented in Table 2. The distribution of errors by complete critical error type is presented in Table 3. As can be seen, all three types of morpheme and syllable exchanges were observed in all three conditions. For the analysis, one factor (WORD TYPE) with three levels was included. Two dependent variables, the number of elicited complete critical errors and the number of all non-critical errors, were measured. Included in the set of non-critical errors are (sub)segmental between-word errors, within-word contextual errors, and non-contextual errors. The mean number of complete critical and non-critical errors were calculated for each participant and for each word pair in a condition. Across conditions, participants made an average of 2.2 critical errors and 11.2 noncritical errors. Each participant produced an average of 1.0 critical error and 3.9 non-critical errors in the free stem condition, 0.98 critical errors and 3.6 non-critical errors in the bound root condition, and 0.2 critical and 3.6 non-critical errors in the control condition. The number of complete critical errors produced for each word pair is given in the Appendix. When examining the distribution of non-critical errors, no main effect of word type was observed by subjects, Friedman: w2(2, N ¼ 42) ¼ 2.25, p 5 .32, or items, Kruskal–Wallis: w2(2, N ¼ 11) ¼ 0.27, p 5 .87. When TABLE 2 Error distribution by condition Error type Free Bound Simple Total Non-contextual errors Contextual within-word errors Contextual between-word errors (Sub)segmental errors All Non-critical errors Critical errors Complete Incomplete 150 6 132 3 122 1 404 10 11 167 18 153 29 152 58 472 43 23 41 7 9 2 93 32 MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 351 TABLE 3 Total number of complete critical speech errors per condition divided into error types Error type Prefix spreads recharge displace Õ recharge replace Stem spreads disorder recover Õ discover recover Total exchanges recharge displace Õ discharge replace Total errors Free stem Bound root Control 12 4 2 12 15 5 19 22 2 43 41 9 examining the distribution of critical errors, a difference between the two prefixed conditions and the control condition was found. The main effect of word type for critical errors was reliable across subjects, Friedman: w2(2, N ¼ 42) ¼ 20.302, p 5 .01, and items, Kruskal–Wallis: w2(2, N ¼ 11) ¼ 9.936, p 5 .01. The distribution of critical errors compared with noncritical errors across the three experimental conditions was also significantly different, w2(2) ¼ 18.99, p 5 .001. Thus, we see that participants produced more critical errors when naming prefixed words than control words but an equal number of non-critical errors across conditions, as predicted. The crucial question under investigation is whether the bound root condition, which consisted of words with a relatively low rate of semantic relatedness, patterned with the free stem condition or the control condition. Pairwise comparisons reveal no difference in the distribution of critical errors between the free stem and bound root conditions by subjects, Wilcoxon Sign: z ¼ 0.043, p ¼ .96, and items, Mann–Whitney U: z ¼ 0.562, p ¼ .61.6 In contrast, significantly more critical errors were produced in the free stem condition, by subjects, Wilcoxon Sign: z ¼ 4.1, p 5 .01, and items, Mann–Whitney U: z ¼ 3.08, p 5 .01, compared with the control condition. Likewise, more critical errors were produced in the bound root condition, by subjects, Wilcoxon Sign: z ¼ 3.42, p 5 .01, and items, Mann–Whitney U: z ¼ 2.27, p 5 .05, compared with the control condition. Since the two prefixed conditions differed significantly in semantic relatedness but not in the number of critical errors produced, it is unlikely that semantics played a large role in inducing the errors. However, there was a great deal of variation between items both in the elicited ratings and 6 Furthermore, there were no significant differences in the type of critical errors produced between the free stem and bound root conditions, Mann–Whitney U: zs 5 1; comparisons within stem spreading errors, prefix spreading errors and total exchanges revealed no significant differences between the two complex word types. 352 MELINGER in the number of errors produced. As a result, it is possible that items with higher semantic ratings produced the most errors. If true, then there should be a significant correlation between the semantic ratings of word pairs and the number of critical errors produced. There are two ways in which this possibility can be investigated. First, the number of critical errors produced by a presentation pair can be correlated with the averaged similarity rating for that pair. Second, rather than averaging the similarity ratings for the two related word pairs together, the ratings for each individual pair of related words can be used for the correlation. Correlating error rates with the semantic ratings from individual related word pairs gives the greatest sensitivity to differences in semantic relatedness ratings. However, it only allows inspection of a subset of the error data, namely the stem exchange errors. Since prefix spreading errors and total exchanges involve the stems/roots from both words, this measure of semantic similarity is inappropriate for them. But, averaging across the individual rating pairs to produce a mean rating for each presentation pair allows one to test whether semantic ratings correlate with the production of all three types of critical errors. Since the interesting comparison is between the two prefixed conditions, only the two prefixed conditions were included in the correlation. The scatter plot in Figure 2 shows the relationship between error rate and averaged semantic rating. Figure 2 shows the relationship between the average semantic rating a presentation word pair received in the rating study and the number of critical errors produced in the WOC task for the bound root and free stem words. As can be seen, there is no correlation between semantic rating and the number of morphological errors produced, Spearman’s Correlation: r ¼ .263, p 5 .237. The lack of a significant correlation suggests that error rates were not influenced by the semantic relationship between a presentation word and its intruding word but rather were due to structural characteristics of the presentation words. The correlation with the averaged similarity rating for each presentation pair allows all critical error types, stem spreads, prefix spreads and total exchanges, to be investigated. However, it loses sensitivity to the specific degree of semantic relatedness of a single word pair, especially when the two rating word pairs that compose a presentation pair receive very different rating scores. To illustrate, consider the presentation pair insist conclude. The two rating pairs for this item received quite different similarity ratings; insist consist received a rating of 2.2 while include conclude was rated much higher, 4.2. This difference in the semantic relatedness within the presentation pair predicts that more errors should be produced involving the root {-clude} than involving the root {-sist}. Since the correlation with averaged semantic ratings cannot address this prediction, a second correlation of semantic relatedness with only the stem MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 353 Figure 2. Distribution of critical errors and semantic ratings for the bound root and free stem word pairs. spreading errors was conducted. Again, only the bound root and free stem words were considered. Stem spreading errors were identified as involving one of the two available stems and the number of errors was correlated with the semantic rating for that stem pair. The scatter plot in Figure 3 shows the relationship between stem error rates and semantic ratings. Again, no correlation was found between number of stem spreading errors and the semantic relatedness of the individual rating pairs, Spearman’s Correlation: r ¼ .241, p 5 .115. As Figure 3 shows, most words produced between zero and two stem spread errors, irrespective of their semantic rating. There are, however, two clear outliers identified. Two word pairs, realign misalign and implode explode, received both high semantic ratings and a disproportionate number of stem spreading errors (their presentation pair counterparts, misassign reassign and expose impose produced stem spreading errors within the normal range). Thus, these two words seem to belie the absence of a correlation between semantic relatedness and error production. However, these two word pairs have another interesting feature; the presentation pairs misassign realign and expose implode have more phoneme overlap than average, three and two segments, respectively. Although this factor was matched across condi- 354 MELINGER Figure 3. Distribution of stem spreading errors and semantic ratings for the bound root and free stem word pairs. tions, there was still some variation between items. These two examples suggest that phoneme overlap may also be a factor that partly predicts the occurrence of critical errors. To ensure that overlap is not a strong predictor of critical error production, a correlation between number of errors produced in all three conditions and segmental overlap was also conducted. This correlation produced a very weak, marginally significant correlation, Spearman’s Correlation: r ¼ .317, p 5 .073. Thus, it seems that phoneme overlap is very weakly correlated with critical error production. While it cannot explain the differences found in critical error production between the two complex conditions and the control condition, it may provide a partial explanation for why certain item sets produced more errors than others. These results clearly indicate that morphological elements are more likely to be exchanged or be substituted in word production than other linguistic units such as syllables. They also suggest that both free stem and bound root words are represented in a similar manner in the mental lexicon; specifically, they are represented either as morphologically decomposed, with each morpheme represented independently, or as whole word forms which include internal morphological structure. If complex MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 355 words are represented as internally structured whole word forms, then this suggests that morpheme boundaries are more salient or accessible to the language production system than syllable boundaries. As predicted, more critical errors were produced in the free stem condition than in the control condition. Likewise, more critical errors were produced in the bound root condition than in the control condition. An equal number of critical errors were produced in the free stem and bound root conditions and there was no correlation between semantic relatedness ratings and the production of critical errors. The difference in error rates for prefixed words compared with control words lends support to models of the mental lexicon that include morphological structure in the representations of derivationally complex words. Since the intended utterances for each error in this study are known, as they were prescribed by the task, alternative analyses which classify morpheme errors together with other word blends between simplex words are ruled out. Also, since the word pairs in the control conditions produced significantly fewer errors, the errors cannot be explained by referring to other levels of linguistic representation such as the syllable. The difference in the number of observed morphological errors in complex words compared with syllable errors in the control words, found in an experimental setting, suggests that the paucity of derivational errors in naturally occurring data may be due to the rarity of clauses containing a derivational affix and two stems, both of which can be concatenated with the affix to produce an actual word. The differences in critical error rates can also not be easily attributed to differences in the intruder word frequencies or string length. For all material differences reported, the free stem condition stood out as different from the other two conditions; the words in this condition were longer and the intruders were less frequent than in either the bound root or control conditions. If differences in critical error rates were attributable to these differences, one would have expected the bound root words to pattern with the control words rather than the free stem words. This is not the observed pattern in error rates. Furthermore, a difference in string length between conditions should have affected the number of non-critical errors as well as critical errors but no difference in non-critical errors was observed. Thus, it is unlikely that these differences had a significant influence on the production of errors in this study. Another difference between the two complex conditions and the control condition in the present study is that the presentation words and intruders consistently are from the same part of speech in the two complex conditions but not in the control condition. In naturally occurring errors, words tend to exchange with words of the same part of speech. However, since naturally occurring errors are always produced in context, it is unclear if a grammatical class constraint operates during lexical access or 356 MELINGER at the stage where words are associated to syntactic positions. If it applies during lexical access, this could be another possible explanation for the present difference in critical error rates. It does not appear, however, that part of speech plays a large role in predicting critical errors in the control condition. Several of the presentation word pairs do not have intruders that are from different grammatical classes yet they still do not produce any critical errors. Similarly, several word pairs that do have intruders from a different grammatical class do produce errors. Thus, while this possible explanation cannot be ruled out completely, the data do not conform neatly to the predictions that arise. The current findings for prefixed words replicate Pillon’s (1998) findings for French suffixed words. One striking difference, however, is that English speakers produced a much higher rate of non-contextual errors compared with the French speakers. Pillon reported that participants rarely failed to reverse the order of words for experimental trials and also rarely produced incomplete responses. In contrast, 30% of responses from English participants included some sort of non-contextual error. Since the rates of non-contextual errors were constant across conditions, one cannot attribute the high rates to the manipulation of word type. Rather, it is more likely due to the differences in the orthographic characteristics of English compared with French. French has a comparatively transparent mapping from orthography to phonology; thus reading times may be faster and less error prone. In contrast, the mapping from orthography to phonology in English is notoriously complex and arbitrary; thus reading times are longer and more error prone. Since the presentation time used in this study was the same as the presentation time used in Pillon’s study, it is possible that some participants were simply unable to fully prepare their response in the time allowed. GENERAL DISCUSSION The results from the present study demonstrate that morphological errors are more frequent than phonological or syllable level errors in laboratoryinduced settings. This replicates Pillon’s (1998) results using prefixed words derived both from free stems and bound roots and it provides further support for a morphological interpretation of naturally occurring morpheme errors. When ordering uncertainty was introduced to the naming task, higher levels of morpheme ordering errors were observed than syllable ordering errors. In agreement with Pillon, this finding suggests that morphemes and morpheme boundaries make up part of a word’s form specification and that these words are treated as complex by the production system in the course of phonetic and/or phonological spell-out. MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 357 More interestingly, these results show that prefixed words derived from free stems and bound roots are equally likely to produce a morpheme exchange error. This finding emerged despite the differing degrees of semantic relatedness in the two prefixed conditions and the differing degrees of semantic relatedness within the morphological families. This equivalency in error frequency suggests that the lexical representations of free stem and bound root words are not qualitatively different. The lack of a significant correlation between presentation pairs’ semantic relatedness ratings and error frequency suggests that errors are due to morpheme misordering and not to the inappropriate selection of an intruder. Taken together, these results support the inclusion of morpheme boundaries in the lexical representation of complex words. This finding has several implications for theories of morphology and models of speech production. Theories of morphology: How are words represented? Morphological theories differ with respect to whether morpheme boundaries are explicitly included in the representations of complex words. Current trends eschew defining the lexicon as the repository of morphemes, preferring to view the lexicon as consisting only of stems or words. Bound morphemes such as affixes have no explicit representation in most current models of the lexicon but rather are expressed by word formation generalisations in a morphological component where knowledge of morphology is expressed. While this is a popular way of conceiving of morphology, there is a great deal of variation between theories in how word formation rules (WFR) are related to the lexical items they are generalising over. In one of the early proposals for WFR use, Aronoff (1976) states that his WFR serve to analyse existing words into their component parts. Thus, in his model, a word’s lexical representation can include morphological structure whether it was created by a WFR or simply analysed by it. Another early proposal for extracting morphology out of the lexicon came from Jackendoff (1975). Jackendoff proposed that morphological relations are expressed by Redundancy Rules (RR). RR are not processes that create words but rather statements that express relationships between words. As a result, Jackendoff’s (1975) model does not acknowledge that words have internal structure. Each word is represented as a single indivisible unit in the lexicon; RR express relations amongst words, but no structure is imposed on words. More recently, Anderson (1992) and Bochner (1993) have extended Jackendoff’s view of an unstructured lexical representation of complex words to include both inflectional and 358 MELINGER derivational morphology using WFR or RR, respectively (but see Carstairs-McCarthy (1992) for several arguments against morpheme-less theories of morphology). The obvious implication of unstructured lexical representations is that they do not predict the type of morpheme errors observed in the current study. The present data clearly suggest that morphemes are treated as units by the speech production system. They can only be treated as component units if lexical representations include information about the morphological make-up of words. Thus, these results argue strongly for the inclusion of morphological structure within the representations of complex words. If we adhere to the view that the lexicon consists of words, not morphemes, these results are most compatible with a model which conceives of WFR as an active part of the lexicon, functioning both to produce new words and to analyse existing words into their constituent parts, and not simply generalisations over an unaltered lexicon. Speech production The mechanism underlying the observed exchange errors was outlined quite explicitly in Pillon’s (1998) general discussion. Specifically, she suggests that when participants view a presentation pair, the words are in the opposite order of how they eventually will have to be named. At this early stage of preparing the response, the morphemes of the first word should have higher levels of activation than the morphemes of the second word. When the response cue informs participants to reverse the order of the words, the activation levels of the words may not adjust quickly enough to avoid misorderings. Thus, morphemes are anticipated, perseverated, or fully exchanged. The present results are best accounted for by models of speech production which include an explicit level of representation for morphemes (Dell, 1986; Levelt, 1989). Production models which do not include an explicit level of morphological representation are in no better position to explain the data than morphological theories which do not include morphological structure in the lexical representations of complex words. The observation that morphological errors are independent of semantic relatedness supports a clear separation between the levels of semantic and morphophonological processing rather than an interaction between them. Discrete speech production models with a clear division between these levels of processing (Levelt, 1989; Levelt, Roelofs, & Meyer, 1999) are better positioned to explain this pattern than models which allow interactions between the levels (Caramazza, 1997; Dell, 1986). Furthermore, the insensitivity of the error mechanism to semantics is also consistent with the separation hypothesis from morphological theory (cf. MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 359 Beard, 1988). The separation hypothesis divides the form of an affix from its function, instantiating them via separate rules. Further evidence both for the inclusion of an explicit level of morphological processing and for the insensitivity of that level of processing to semantic transparency comes from recent reaction time studies in speech production. Specifically, Roelofs & Baayen (2002) conducted a reaction time study using the implicit-priming paradigm to investigate the role of semantics to the production of morphologically complex words. They constructed sets of three words with the same word onset. They varied whether all three words were complex and semantically transparent, or whether two complex and semantically transparent words were combined with one phonologically matched simple word or a phonologically matched morphologically complex word that was semantically opaque, such as in (17). (17) a. input insight inflow b. input insight insect c. input insight invoice They found that semantically opaque words patterned with the semantically transparent words while the morphologically simple words did not. Thus, they interpret their results as additional evidence for a structure of the language production system that strictly separates semantic information from morphophonological information (Levelt et al., 1999). Why don’t we get naturally occurring morpheme errors? Given the frequency of morpheme exchanges observed in this study, as well as Pillon’s (1998) study, the question remains as to why derivational errors are so rare in naturally occurring data. There are several possible explanations. First, there is a strong lexical bias in speech production. Errors resulting in non-words are much less frequent than errors resulting in actual words (as the biased vs. unbiased contrast in Pillon’s study demonstrated). The speech production monitoring system (Levelt, 1989) may catch non-lexical errors more easily than it catches lexical errors. Most syllable and morpheme exchanges result in illegal combinations. As a result, many possible morpheme errors are filtered out in the course of speech production before they are ever uttered. A second possible explanation is that many errors are uncorrected in natural speech. Given the semantic relationship between words which share a free stem (or in many cases even a bound root), the error may not be clearly noticeable. For example, if a speaker says ‘‘I need to react 360 MELINGER quickly’’ when the intended utterance was ‘‘I need to act quickly’’, the error may go uncorrected and unnoticed. This possibility exists for many prefix þ stem combinations, especially since syntactic violations, which clearly signal an error, do not result from errors involving English prefixes since they do not change the target word’s part of speech. This is crucially different from suffix exchanges, such as ‘‘I arrived at the hotel’’ Õ ‘‘I arrival at the hotel’’, in which the error is clearly noticeable. Thus, the frequency of morpheme errors in naturally occurring data may, in fact, be somewhat underestimated, especially for prefixed words. A third possibility returns to the original issue of interpretablity. Since many morphological errors can alternatively be interpreted as phonological or semantic errors, many observed errors may be recorded but not classified as morpheme errors. This would also lead to an underestimation of morpheme errors in naturally occurring data. In summary, the current results clearly demonstrate that morpheme errors do occur more frequently than phonological errors involving syllables, as was previously shown by Pillon (1998). More importantly, the data show that the observed errors are purely morphological and not semantic. Bound root and free stem words produce the same number of errors suggesting that their lexical representations are not qualitatively different. These results are best accounted for by models of morphology that explicitly include morphological structure in the lexical representation of complex words and models of speech production which explicitly include a level of morphological processing that is insensitive to semantic factors. Manuscript received July 2002 Revised manuscript received November 2002 REFERENCES Anderson, S. (1992). A-Morphous morphology. Cambridge: Cambridge University Press. Aronoff, M. (1976). Word formation in generative grammar. Cambridge, MA: MIT Press. Baars, B., & MacKay, D. (1978). Experimentally eliciting phonetic and sentential speech errors: Methods, implications, and work in progress. Language in Society, 7, 105–109. Baars, B., Matteson, M., & Cruickshank, G. (1985). The induction of syllable exchanges. Unpublished manuscript, Department of Psychology, State University of New York, Stony Brook. Baars, B., & Motley, M. (1976). Spoonerisms as sequencer conflicts: Evidence from artificially elicited errors. American Journal of Psychology, 89, 467–484. Baars, B., Motley, M., & MacKay, D. (1975). Output editing for lexical status in artificially elicited slips of the tongue. Journal of Verbal Learning and Verbal Behavior, 14, 382–391. Beard, R. (1988). On the separation of derivation from morphology: Toward a Lexeme/ morpheme-based morphology. Quaderni di Semantica, 9, 3–59. MORPHOLOGICAL STRUCTURE OF PREFIXED WORD 361 Bochner, H. (1993). Simplicity in generative morphology. Publications in Language Sciences, Vol. 37. New York: Mouton de Gruyter. Bybee, J., & Slobin, D. (1982). Rules and schemas in the development and use of English past tense. Language, 59, 251–270. Caramazza, A. (1997). How many levels of processing are there in lexical access? Cognitive Neuropsychology, 14, 177–208. Carstairs-McCarthy, A. (1992). Morphology without word-internal constituents: A review of Stephen R. Anderson’s A-Morphous Morphology. In G. E. Booij & J. van Marle (Eds.), Yearbook of morphology 1992 (pp. 209–233). Dordrecht: Kluwer Academic Publishers. Cutler, A. (1980). Errors of stress and intonation. In V. Fromkin (Ed.), Errors in linguistic performance: Slips of the tongue, ear, pen, and hand (pp. 67–80). New York: Academic Press. Dell, G. (1986). A spreading activation theory of retrieval in sentence production. Psychological Review, 93, 283–321. Dell, G., & Reich, P.A. (1977). A model of slips of the tongue. In R.J. DiPietro & E.L. Blansitt, Jr. (Eds.), LACUS Forum 3 (pp. 448–455). Columbia, SC: Hornbeam Press. Dell, G., & Reich, P.A. (1980). Toward a unified model of slips of the tongue. In V. Fromkin (Ed.), Errors in linguistic performance: Slips of the tongue, ear, pen, and hand (pp. 273– 286). New York: Academic Press. Fromkin, V. (1971). The nonanomalous nature of anomalous utterances. Language, 47, 27–52. Fromkin, V. (1973). Speech errors as linguistic evidence. The Hague: Mouton. Garrett, M.F. (1980). Levels of processing in sentence production. In B. Butterworth (Ed.), Language production. London: Academic Press. Garrett, M.F. (1982). Production of speech: Observations from normal and pathological language use. In A.W. Ellis (Ed.), Normality and pathology in cognitive functions (pp. 19– 76). London: Academic Press. Jackendoff, R. (1975). Morphological and semantic regularities in the lexicon. Language, 51, 637–671. Jescheniak, J.D., & Levelt, W.J.M. (1994). Word frequency effects in speech production: Retrieval of syntactic information and of phonological form. Journal of Experimental Psychology: Learning, Memory, & Cognition, 20, 824–843. Levelt, W.J.M. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press. Levelt, W.J.M., Roelofs, A., & Meyer, A.S. (1999). A theory of lexical access in speech production. Behavioral and Brain Sciences, 22, 1–75. MacKay, D. (1976). On the retrieval and lexical structure of verbs. Journal of Verbal Learning and Verbal Behavior, 15, 169–182. MacKay, D. (1978). Derivational rules and the internal lexicon. Journal of Verbal Learning and Verbal Behavior, 17, 61–71. MacKay, D. (1980). Speech errors: Retrospect and prospect. In V.A. Fromkin (Ed.), Errors in linguistic performance: Slips of the tongue, ear, pen, and hand (pp. 319–332). New York: Academic Press. Neely, J. (1976). Semantic priming and retrieval from lexical memory: Evidence for facilitatory and inhibitory processes. Memory and Cognition, 4, 648–654. Pillon, A. (1998). Morpheme units in speech production: Evidence from laboratory-induced verbal slips. Language and Cognitive Processes, 13, 465–498. Plaut, D., & Gonnerman, L. (2000). Are non-semantic morphological effects incompatible with a distributed connectionist approach to lexical processing? Language and Cognitive Processes, 15, 1–19. Roelofs, A. (1996). Serial order in planning the production of successive morphemes of a word. Journal of Memory and Language, 35, 854–876. 362 MELINGER Roelofs, A. & Baayen, H. (2002). Morphology by itself in planning the production of spoken words. Psychonomics Bulletin and Review, 9, 132–137. Rueckl, J. G., Mikolinski, M., Raveh, M., Miner, C., & Mars, F. (1997). Morphological priming, fragment completion, and connectionist networks. Journal of Memory and Language, 36, 382–405. Stemberger, J. (1985). Bound morpheme loss errors in normal and agrammatic speech: One mechanism or two? Brain and Language, 25, 246–256. Stemberger, J.P., & MacWhinney, B. (1986). Frequency and the lexical storage of regularly inflected forms. Memory and Cognition, 14, 17–26. Swinney, D., Onifer, W., Prather, P., & Hirshkowitz, M. (1979). Semantic facilitation across sensory modalities in the processing of individual words and sentences. Memory and Cognition, 7, 159–165. Tanenhaus, M., Flanigan, H., & Seidenberg, M. (1980). Orthographic and phonological activation in auditory and visual word recognition. Memory and Cognition, 8, 513–520. Zwitserlood, P., Bölte, J., & Dohmes, P. (2000). Morphological effects on speech production; Evidence from picture naming. Language and Cognitive Processes, 15, 563–591. APPENDIX Free stem words Recharge Displace Misnamed Unused Refold Enact Unordered Discounted Misassign Realign Untitled Subconscious Reappear Disunite Misprint Prejudge Enclose Distrust Reload Unwind Disorder Recover a No. of errors Bound root words No. of errors 5 4 5 1 6 0 4 5 2 3 8 Insist Conclude Detain Observe Concede Profess Infer Transduce Expose Implode Perturb Dissuade Decrease Inflate Induce Reject Erect Corrupt Dispose Propel Remit Advise 1 1 0 6 13 2 5 2 4 0 7 Control words Miller Parcel Mugger Baffleda Mellow Furnace Socket Pallid Rattle Morbid Loyal Tower Scanty Pedal Defile Probate Certify Forgery Setter Fashion Briar Tidal Number of errors 0 0 0 0 0 3 1 1 1 0 3 The intruding word bagger is an American term that refers to a person at a grocery store who places your purchased items into a plastic or paper bag.
© Copyright 2025 Paperzz