Partial Knowledge of Word Meanings

Journal of Experimental Psychology: General 1991, Vol. 120, No. 2, 190-202

Copyright 1991 by the American Psychological Association, Inc. 0096-3445/91/S3.00

Partial Knowledge of Word Meanings Francis T. Durso and Wendelyn J. Shore University of Oklahoma Partial knowledge of word meanings was investigated in 7 experiments. For each S, words were classified depending on whether they could be defined (known), checked as familiar (frontier), or mistakenly considered to be nonwords (unknown). Even when Ss denied a word was a part of the language, they often were able to distinguish between correct and incorrect uses of the word. Explanations based on materials, retrieval environments, and task demands could not account for the effects. Ss often had meaningful information about even unknown words (including definitional and contextual meaning), although they were unable, or unwilling, to use it. Although frontier words surpass unknown words in their access to information relevant to familiarity and lexicality judgments, both unknown words and frontier words can supply comparable information about meahing. The information used to judge familiarity or lexicality does not appear to be the same information used to determine the appropriate use of the word.

The importance of partial knowledge has been recognized for a number of years by educators and researchers interested in the representation of expertise (e.g., Chase & Simon, 1973), concept formation (e.g., Hayes-Roth, 1977; Hintzman, 1986), the development of cognition (e.g., Wilkinson, 1982), the learning of skills (e.g., Anderson, 1982), the development of word meaning (Clark, 1973; Nelson, 1982), and the acquisition of vocabulary (e.g., O'Connor, 1940). Despite the efforts of these scholars, most areas of cognitive science implicitly assume that a model of partial knowledge will be derived from models of complete knowledge; similarly, results based on studies using familiar stimuli are assumed to apply, with perhaps some minor modification, to less familiar materials. Such simplifying assumptions are not unique to cognitive psychology: For example, in psychometrics, the knowledgeor-random-guessing principle of latent trait theory explicitly assumes that the subject either knows the answer or guesses randomly (Lord & Novick, 1968). An interesting phenomenon has emerged in a variety of studies that is not easily interpreted by applying existing models of complete knowledge. People perform surprisingly well on tasks when they claim to have partial or no knowledge of the target. As examples, in a study investigating vocabulary acquisition, Loewenthal (1971) noted that subjects performed well on words they did not know; Koriat and Lieblich (1974), in a follow-up to Brown and McNeill's (1966) classic tip-of-

the-tongue (TOT) study, showed that subjects in a "don't know" state could, like those in a TOT state, supply partial information about the structure of the target word; Eysenck (1979) showed in a feeling-of-knowing (FOK) study that the subject could make semantic differential judgments (e.g., good word or bad word) on words unknown to the subject; and finally, Shore and Durso (1990) reported subjects performed well above chance on a forced-choice sentence-decision task, even when subjects denied that the target word was part of the English language. In this article, we investigate the ability of people to make judgments involving words about which they claim to have partial or no knowledge. A direct investigation of partial knowledge of word meanings would have implications for models of lexical memory and theories of the development of word meaning. In addition to supplying direction for modification of models that do not easily accommodate partial knowledge, the findings have obvious implications for instruction. In such an effort, it is important to consider partial knowledge of word meanings from two perspectives. One perspective is the growth in knowledge that presumably occurs as the meaning of a word is learned: What information is available from familiar words that is not available from unfamiliar ones? A second perspective is not on semantic acquisition, but on the metamemorial processes that subjects use in deciding whether the search for meaning is warranted. We begin by reviewing efforts to articulate the levels of word knowledge through which a letter string passes during acquisition. We then consider some of the work on metamemorial phenomena, including research on the TOT state and FOK judgments, that bears on subjects' ability to retrieve information about words that they claim are unknown.

This work was supported by grants from the University of Oklahoma Associates Fund. We would like to express our appreciation to Dirk LeFlore, Cornel Grantham, Laura Young, and Lynne Hill for their help in data collection. Special thanks to Karen Bailey for her efforts in the freeassociation experiment. Thanks also to Rich Reardon, Scott Gronlund, Karen Bailey, and Penny Kariznowi for their comments on earlier versions of this article. Correspondence concerning this article should be addressed to Francis T. Durso, Department of Psychology, 455 West Lindsey Street, University of Oklahoma, Norman, Oklahoma 73019-0535. Electronic mail may be sent to [email protected].

Levels of Word Knowledge At one extreme of word knowledge are unknown words. At this level, subjects would deny that they have seen the word before (Dale, 1965). These are words that subjects are presum-

190

191

PARTIAL KNOWLEDGE

ably unable to distinguish from nonwords if asked to do so (Curtis, 1987). As others (Drum, 1983; O'Connor, 1940) have pointed out, words at this level presumably bring to mind already known words similar in appearance or sound. On a vocabulary test, subjects with this level of word knowledge may choose a synonym of the sound-alike-look-alike associate. At the other extreme are known words. These are words that subjects claim they know and for which they can, in fact, provide a specific definition that is not tied to any particular context (Dale, 1965; Drum, 1983; O'Connor, 1940). These words are likely part of the subjects' expressive vocabulary. In addition to these two extremes of word knowledge, there exists at least one intermediate level in all characterizations of word meanings. This intermediate level has been less well operationalized, but is the level that can most clearly be classified as partial knowledge. Words at this level can be recognized as real words, but subjects deny any knowledge of their meaning and cannot define them. Subjects may be able to place the word in the correct general syntactic or pragmatic context, but have no real idea of the word's meaning (Drum, 1983). Trembly (1966) labeled words at this stage as frontier words because they exist "in the frontier region between the point where every word is known and the point where no words are known" (p. 229). We will adopt his label for words that are judged as familiar but that the subject cannot define. Although schemes for the classification of word knowledge have often included additional levels beyond the unknownfrontier-known scheme, there is less agreement on the characteristics of these levels. For instance, O'Connor (1940) includes two additional levels. At one level, the word is tied closely to an episodic context that the subject remembers; at another level, the target's meaning is confused with its antonym. Because scholars disagree on the number and characteristics of stages beyond a tripartite scheme, and because very little empirical research has been done investigating even the simple three-part classification, the research reported here restricts itself to comparisons of unknown, frontier, and known words. Although researchers (e.g., Curtis, 1987; Drum & Konopak, 1987) have acknowledged the existence of partial knowledge of word meanings, much of the empirical work on word knowledge has been concerned with only known and unknown words. The research that has been done on partial word knowledge (e.g., Jenkins, Stein, & Wysocki, 1984; Loewenthal, 1971; Nagy, Herman, & Anderson, 1985) has been criticized for failing to provide an accurate assessment of this level (Drum & Konopak, 1987). For instance, some simply assume that selecting uncommon words as stimuli insures that subjects know nothing about the target words (e.g., Jenkins et al, 1984). When prechecks are used to assess subjects' preexperimental vocabulary knowledge, the prechecks usually do not extract information that confirms subjects' self-assessment of word knowledge. For example, Nagy et al.'s (Eysenck, 1979; Loewenthal, 1971) subjects supplied no definitions; subjects simply made yes-no decisions as to whether or not they knew the meaning of the word. This method "cannot measure partial knowledge or incremental gains in knowledge" (Drum & Konopak, 1987, p. 79). In fact, Nagy et al. simply assumed that words at the frontier level

would be marked as known by these subjects. Finally, Curtis (1987) makes the point that to assess the unknown level requires a precheck that distinguishes between words subjects identify as nonwords and words they identify as real words. Shore and Durso (1990) attempted to meet the criticisms leveled against earlier work by asking subjects to complete a precheck derived from the theoretical word levels proposed by Dale (1965). From the precheck, Shore and Durso classified words into one of three levels. Known words were those the subject could correctly define or could use in a sentence. Frontier words were words that the subject checked as vaguely familiar, but about which they could give no meaningful information. Unknown words were not only unfamiliar, they were words the subject claimed were nonword letter strings and therefore not part of the language. Although the focus of Shore and Durso (1990) was on the effectiveness of instruction as a function of word level, of importance here is performance of their control subjects, those who received no instructional intervention. After the assessment of word level, subjects participated in a forced-choice sentence decision task. Half of the subjects chose between a sentence that used the word correctly (e.g., The dowager has her dead husband to thank for her great wealth.) and a sentence that violated the general semantic constraints of the word (e.g., The circus performer could juggle the dowager and two bowling pins.). The other half of the subjects chose between a sentence that used the word correctly and a sentence that violated specific details of the word's meaning (e.g., The dowager and her husband recently celebrated their fiftieth wedding anniversary.). Subjects were able to choose the correct sentence more often as word-level knowledge increased and as the violations of the incorrect sentence became more severe. Surprisingly, subjects were significantly above chance in choosing the sentence that used an unknown word correctly over one that violated general constraints of the unknown word: Subjects chose correctly 70% of the time, even though they had denied the word was a legal English word. In summary, conceptualizations of word levels are unanimous in the inclusion of at least three levels: an unknown level, an intermediate-vaguely familiar-frontier level, and a known level. In addition, the little empirical work that does exist suggests that subjects seem to have some knowledge about even unknown words. Knowing more than one thinks is not, however, an uninvestigated phenomenon.

Metamemory In any investigation of metamemory, the subjects' assessment of what they know is compared with performance on another task that can be objectively evaluated. This characterizes a number of paradigms that most recently would be classified under the FOK rubric (Hart, 1965). In a typical FOK experiment, subjects are asked to recall a piece of information and then to make judgments of the recognizability of information they could not recall. Finally, a recognition test is given. Performance on the recognition test is correlated with the subjects' FOK judgments, yielding the dependent measure of interest.

192

FRANCIS T. DURSO AND WENDELYN J. SHORE

Items that are given a positive FOK judgment are recognized more often than those given a negative FOK judgment (Blake, 1973; Hart, 1967; Nelson, Leonesio, Shimamura, Landwehr.&Narens, 1982; Nelson &Narens, 1980;Schacter, 1983). Typically, the size of these correlations is moderate. In fact, normative item difficulty is a better predictor of recognition performance than is the individual's FOK judgment (Nelson, Leonesio, & Landwehr, 1986). The correlations tend to be modest because subjects often seemed to possess information, as indicated on the recognition test, that they denied having, as indicated on the FOK judgments. Even the earliest work (Hart, 1965) indicated that when subjects "very strongly" denied knowing an answer, they still scored slightly above chance. When the denial was only slightly less strong they scored significantly above chance. In fact, these latter scores were not statistically different from items given a very strong positive FOK rating. Another instance of successfully monitoring memory when information is, for the time being, inaccessible is the empirical work investigating James's (1890) TOT phenomenon. Much of the work that has investigated memory monitoring has asked subjects to retrieve a lexical entry given some retrieval cue, such as a word's definition, a person's face, or an associated fact (Blake, 1973; Brown & McNeill, 1966; Koriat & Lieblich, 1974; Lovelace, 1987; Yarmey, 1973). Of interest in this research was the partial information to which subjects have access when recall fails, but the subjective feeling is one of possessing some knowledge. The sought-for-target in memory has not always been a simple lexical entry. For example, in the Eysenck (1979; Shore & Durso, 1990) work discussed earlier, subjects were presented with uncommon words and then attempted to ascertain the semantic attributes available for words that the subjects "definitely did not" know the "precise meaning of the word." Even for these extreme responses in Eysenck's study, subjects were able to make judgments about where the word fell on a semantic differential. Of course, denying knowledge of the precise meaning of a word may still allow subjects to have knowledge of some meanings; that is, the word may be more familiar to the subject than Eysenck assumed. However, this argument is more difficult to apply to the Shore and Durso results in which subjects performed above chance for words they denied were part of the language. Because above-chance performance for items judged to be nonwords has been reported only once in the literature (Shore & Durso, 1990), we felt it prudent to begin with a replication. A successful replication of Shore and Durso would show that subjects perform better as the level of word knowledge increases; subjects can easily pick between a correct use of the word and a use that violates general constraints; and most importantly, that this ability to pick the correct sentence obtains even when subjects deny that the target word is a part of the English language.

Experiment 1 The level of word knowledge of each word—known, frontier, and unknown—was assessed for each subject. Half of the

subjects then chose between a sentence that used the word correctly and one that violated general constraints; half of the subjects chose between a correct sentence and one that violated specific, detailed information. The assessment of word knowledge, followed by a binary decision between stimuli that use the word correctly and incorrectly, characterizes in general the procedure used in each study reported in this article.

Method Subjects Subjects were 60 volunteers from the introductory psychology subject pool at the University of Oklahoma who participated in this experiment as part of a research familiarization requirement. All subjects were native English speakers. They were randomly assigned to the between-subjects variable of sentence type (constraint-only vs. specific detail). Random assignment was effective in equalizing verbal intelligence as measured by the vocabulary subscale of the ShipleyHartford Intelligence Test (Sines, 1958): Mean for those seeing constraint-only sentences was 28 and for those seeing specific, detail sentences, 29 (maximum of 40). Word level (unknown, frontier, and known) was a within-subjects variable.

Materials Precheck. A precheck was used to determine the level of word knowledge for each word for each subject. The 60 words (see Appendix A) used by Shore and Durso (1990), found to span the range of word knowledge, were used. In addition, 15 pronounceable nonwords (e.g., edarthic) were included on the precheck. The nonwords were constructed by the experimenters and followed the rules of English orthography. They were used to obtain a purer measure of the unknown word level (Curtis, 1987). Sentence pairs. Three sentences that used the target word were constructed for each of the 60 words (see Appendix B). All three sentences contained a syntactically correct usage of the target word. One sentence used the word correctly (e.g., The dowager has her dead husband to thank for her great wealth.), and two sentences used the word incorrectly, either by violating a general constraint or by violating a more specific detail of the word's meaning. Thus, whether subjects were in the constraint-only or the specific-detail sentence conditions depended on the type of incorrect sentence with which the correct sentence was paired. Incorrect sentences in the constraint-only condition contained an anomalous use of the target word (e.g., The circus performer could juggle the dowager and two bowling pins.). Choosing between the correct and incorrect sentences in this condition required knowledge of only the most general constraints of the word, such as knowing that a dowager is something animate, or that a dowager is a person and therefore something that cannot be juggled. In the specific-detail condition, more complete knowledge of the meaning of the target word was required to choose correctly between the correct and incorrect sentences. Here, the incorrect sentences violated the more specific meaning of the target word (e.g., The dowager and her husband recently celebrated their fiftieth wedding anniversary.). In this condition, it was not enough for subjects to know that dowager refers to something animate or to a person. Rather, subjects had to know that a dowager is a kind of widow to choose the correct sentence. Simply knowing the general constraints of the target word was insufficient. For half of the target words, the correct sentence was the first member of the pair, whereas for the other half, the incorrect sentence

193

PARTIAL KNOWLEDGE was the first member. The sentences were always presented in pairs, but the order of the pairs was randomized across subjects.

Table 1 Mean Proportion Correct and Standard Errors (SEs) on the Sentence Decision Task in Experiment 1

Procedure

Word level

Shipley-Hartford Intelligence Test. On entering the laboratory, subjects completed the vocabulary subsection of the Shipley-Hartford. This test was used to measure subjects' preexperimental vocabulary knowledge. Completion of the Shipley-Hartford was subject-paced. No words from the Shipley-Hartford appeared elsewhere in the experiment. Precheck. Following completion of the Shipley-Hartford, subjects were administered the precheck composed of the 60 target words and 15 nonwords. Progression through the precheck was subject-paced. Order of the 75 items was randomized across subjects. On their first pass through the precheck, subjects were told to write a brief definition or synonym for each of the items whose meaning they knew. They were told that if they were not sure of the meaning of an item, they should leave it blank. On completion of the definition pass, subjects were asked to use any remaining items in a sentence. They were told that if they could not use the word in a good sentence, they should leave it blank. Those words that were correctly defined or used correctly in a sentence by a subject formed the set of known words for that subject. Incorrectly defined words or words used incorrectly in a sentence were discarded. After the first two passes had been completed, subjects were asked to check any of the remaining items with which they were familiar: "Place a check mark to the left of the word if it is familiar to you. In other words, if you've heard or seen the word before, or if it looks familiar to you in any way, place a check next to it, even though you may not have any idea what it means." Those words checked as familiar by a subject formed the set of frontier words for that subject. Subjects were then instructed to make one final pass through the precheck. They were told that there were some nonwords on the list and that their task was to circle any of the remaining items that were real English words. It was stressed that any item left completely unmarked was a letter string that they thought did not exist in the English language as a word. Those words that were left unmarked by a subject formed the set of unknown words for that subject. Circled words were discarded. Thus, the words forming the unknown, frontier, and known conditions varied from subject to subject. On completion of the precheck, subjects spent 10 min on a distractor task in which they located and traced simple geometric forms within more complex geometric arrays. Sentence-decision task. After the 10-min delay, subjects were randomly assigned to either the constraint-only or the specific-detail sentence condition. They were told that they would see pairs of sentences and that one member of the pair was a correct use of the underlined word, whereas the other was an incorrect use. They were assured that one of the sentences would always use the underlined word correctly and that if they were not sure which sentence this was, they should guess. The instructions stressed that they must choose one sentence from the pair. They could not leave any blanks. The sentence-decision task was subject-paced. When it had been completed, subjects were informed as to the purpose of the experiment and thanked for their participation.

Results Proportion correct (see Table 1) on the sentence-decision task was computed at each of the three word levels (unknown, frontier, and known) for each subject. These proportions were submitted to a 2 (sentence type: constraint-only or specific-

Sentence type Constraint only Proportion correct

SE Specific detail Proportion correct

SE

Unknown

Frontier

Known

.63

.79

.97

.035

.031

.008

.56

.65

.94

.049

.036

.013

detail) x 3 (word level: unknown, frontier, and known) multivariate analysis of variance (MANOVA). The MANOVA was used to control for violations of the sphericity assumption that may occur across the repeated factor (Winer, 1971). All tests were at an alpha level of 0.05 or better. The MANOVA revealed two significant main effects: The constraint-only sentences were always easier than the specificdetail sentences, F(l, 58) = 6.74, and the more subjects knew about the word, the more accurate they were, F(2, 57) = 95.27. One-tailed contrasts against chance revealed that subjects always performed above chance in the constraint-only condition, even for their unknown words. These results are consistent with those of Loewenthal (1971) and Shore and Durso (1990). At least under these experimental circumstances, the phenomenon is reproducible, and the results allow for further investigations to determine the limits of the phenomenon and the processes underlying it. One limitation of the phenomenon is apparent from the specific-detail condition: Subjects performed significantly better than chance when the sentences contained their frontier and known words, but not when the sentences contained their unknown words. It appears then that the subjects can make accurate decisions involving the general constraints of a word's meaning but perform at chance when the prerequisite knowledge is that of specific definitional features. This finding is consistent with Nelson, Fehling, and Moore-Glascock's (1979) work on savings effects in episodic memory. Nelson et al. supply evidence that traces of forgotten items contain semantic information about superordinate and subordinate relations, but not information about semantic relations at the same level (e.g., synonyms, antonyms) as the forgotten word. The hypothesis of Nelson et al. about the nature of semantic information in subthreshold episodic memories may also apply here. The specific definitional features unavailable to the unknown words are at the same level of inclusiveness as the unknown word; the constraint information is at a higher level of inclusiveness. Thus, savings of unrecognized words and semantic decisions about unknown words both appear to suggest that general constraint information is present in the memory trace. Each of the remaining experiments focuses on decision tasks that require an understanding of the general constraints of the target word.

194


Experiment 2

Results

Experiment 2 was designed to rule out the possibility that subjects performed above chance because the correct sentences were perceived as better English sentences than were the incorrect sentences, regardless of the presence of the target word. It is possible that when the experimenters created correct sentences for a target word, the construction of the sentences was more sound than the construction of sentences in which the target word did not fit. To investigate this possibility, a blank replaced the target words in the general constraint sentence pairs from Experiment 1. We asked 30 subjects to pick the "better" English sentence. Subjects were equally likely to pick the sentence frame that was correct in Experiment 1 (52%) as they were to pick the incorrect sentence frame from Experiment 1 (48%). Thus, structural differences in the sentence frames alone could not account for the above-chance performance in Experiment 1.

Mean proportion correct on the sentence decision task was again computed at each of the three word levels for each subject. The MANOVA revealed the higher the level of word knowledge, the better subjects performed, F(2, 27) = 62.67, replicating Experiment 1. Contrasts against chance also replicated Experiment 1. Subjects were reliably above chance for known (.96; SEM = .017), frontier (.73; SEM = .033), and most importantly, unknown (.61; SEM = .033) words. Thus, the change in retrieval context between the precheck and the sentence decision task was not sufficient to account for the better-than-chance performance observed in Experiment 1.

Experiment 4 Experiments 4 and 5 attempted to determine the limits of the above-chance performance found in Experiment 1. Experiment 4 assessed the necessity of the contrast provided by the presentation of a correct and incorrect sentence. Subjects made a yes-no decision on individually presented sentences.

Experiment 3 Another explanation of above-chance performance for unknown words relies on the well-established principle that memory performance varies with the richness of the retrieval environment and that memory is better when the retrieval context matches the encoding context (Tulving, 1979; Tulving & Thompson, 1973). On the precheck, when subjects deny the letter string is an English word, the only retrieval cue is the isolated word. During the sentence decision task, the pair of sentences provides a much richer cue for memory retrieval. It is possible that subjects accessed information during the sentence decision, which—if it had been accessed earlier— would have led them to check the word as vaguely familiar or even to define it correctly. Thus, the argument is that subjects do not have knowledge of unknown words, but merely that they incorrectly classified the word as unknown on the precheck because of a retrieval failure. We investigated this possibility by providing subjects with the sentence-pair context during the precheck. Thus, subjects decided whether a word was vaguely familiar, or decided whether the letter string was a word, in the context of the same sentence pairs they would ultimately see in the sentence decision task.

Method Subjects were 30 volunteers from the same pool as those in Experiment 1. No one had participated previously. One subject who did not have words at each of the three levels of word knowledge was eliminated. Mean performance on the Shipley-Hartford was 30. All materials were identical to those used in Experiment 1, except that only the general constraint pairs were used. When making their frontier and unknown word decisions, subjects were given a modified precheck that included the appropriate sentence pair under each item. Subjects were told that they should use the information from these sentences to help them make their familiarity and lexical decisions. In addition, two sentences were constructed for each of the nonwords.

Method Subjects Participants were 30 volunteers from the same pool as the previous experiments. We eliminated 4 subjects who did not have instances of each level of word knowledge. The Shipley-Hartford mean was 28. None had served as subjects in previous experiments.

Materials The word pool and precheck were identical to those used in Experiment 1. Instead of sentence pairs, however, single sentences from the constraint-only condition were used on the sentence decision task. Underneath each sentence the words yes and no appeared. Two alternate forms were constructed, both of which contained 30 correct sentences and 30 incorrect sentences. Assignment of the correct sentence was counterbalanced across subjects so that no subject responded to both the correct and incorrect version of a sentence. By the end of the experiment, each sentence was responded to by an equal number of subjects. Order of the 60 sentences was randomized across subjects.

Procedure After completing the Shipley-Hartford, the precheck, and the distractor task, subjects were given the modified sentence decision task. They were told that they would see a series of sentences one at a time and that their task was to decide whether the underlined word in each sentence was used correctly. If so, they were to circle yes, and if not, they were told to circle no. As before, subjects were instructed to guess if they were uncertain.

Results Mean proportions correct on the sentence decision task were computed as before, and these means were submitted to a one-way MANOVA. As before, the MANOVA revealed a

195

PARTIAL KNOWLEDGE

significant main effect of word level: The more subjects knew about a word, the more accurate they were on the sentence decision task, F(2, 24) = 17.30, with known words (.86; SEM = .041) outperforming frontier words (.63; SEM = .035), which outperformed unknown words (.50; SEM = .049). Analysis of the means, however, reveals an interesting difference between these data and those reported previously. In Experiments 1 and 3, subjects performed significantly above chance on the constraint-only sentences at all word levels. However, when single constraint-only sentences rather than pairs were used, subjects were not significantly above chance on their unknown words. Performance on frontier words remained significantly above chance, as did performance on known words, although accuracy at these higher levels was lower than that previously seen. Thus, for subjects to do more than guess, some contrast must be present at test for their unknown words. This is consistent with Experiment 1 in which very subtle contrasts (specific sentences) did not allow subjects to choose the correct sentence. It is also noteworthy that frontier words continued to be above chance here, as they were with the subtle contrasts of Experiment 1. These findings suggest that subjects' frontier words are not simply more familiar than unknown words, but have more—perhaps different—information. The finding that frontier words will outperform unknown words is, of course, the natural expectation. However, as we will see, performance on unknown words is not always inferior to performance on frontier words.

Experiment 5 Experiment 5 attempted to partition the information contained in the sentence pairs used previously into definitional information and selectional-restriction information (contextual). Subjects picked between a synonym and unrelated word (definitional) or between a legal phrase using the word and an illegal phrase (contextual). When subjects decide whether a word is used correctly, information about the word, information about the sentence, and information about the interaction of the word with the sentence can be used to make the decision. We have (e.g., Experiment 2) ruled out information about the sentence alone, but the information that subjects are using remains unclear. In Experiment 5, we looked at two of the many types of information that might be useful in choosing between correct and incorrect information. When a word appears in a sentence, subjects have access to definitional (meaning) information about the word, and they have access to contextual (selectional-restriction) information associated with the word (Miller, 1978). In previous experiments, subjects in the constraint-only condition may have been able to pick one sentence over the other because they realized from the definitional information that the word did not fit. For example, widow shares many definitional components with dowager, whereas plate shares few features with dowager. Alternatively, subjects might perform above chance because they have access to the contextual constraints associated with words, without having any information about the definitional meaning of the word. For example, dowager can

be modified by friendly but not by the adjective, broken. Thus, a choice between sentences used in the previous studies could have been based on definitional information, selectional restrictions, or both.

Method Subjects Subjects from the same pool were tested until data from 75 volunteers could be used in the experiment. We replaced 6 subjects for whom not all levels of word knowledge were indicated on the precheck. None had previously participated in these studies. ShipleyHartford scores were 29 for each of the three conditions: definitional, contextual, and combined.

Materials Rather than sentence pairs, three types of word pairs were constructed for each target word. Definitional pairs consisted of a synonym of the target word and an unrelated foil. For example, the definitional pair for dowager was widow—plate. The contextual pairs contrasted a context in which the word could appear with a context in which the word could not appear. In the case of dowager, subjects chose between friendly dowager and broken dowager. The constituents of the phrase bore an arbitrary relationship to the target word. For example, there is nothing in the meaning of dowager that necessitates or prohibits the modifier friendly. The third type of word pair was a combination of the first two. This type provided both definitional and contextual information about the word, with subjects choosing between friendly widow and broken plate for the dowager target. All subjects were instructed to pick the member of the pair that was closer in meaning to the target word. Again subjects were forced to choose for each pair. Except for the above changes, methodology was identical to Experiment 1.

Results Mean proportions correct on the word pair decision task appear in Table 2. These means were analyzed by a 3 (word pair type: definitional, contextual, or combination) x 3 (word level: unknown, frontier, and known) MANOVA. Replicating previous results, there was a significant main effect of word level, F(2, 71) = 111.89. This effect was due to known words outperforming either frontier words or unknown words. In fact, there was little indication that frontier words were better than unknown words in this experiment. Thus, in some circumstances, subjects can perform as well on items they denied were part of the language as on items they claimed were familiar. In addition, there was a significant main effect of word pair type, F(2, 72) = 3.33. This effect was due to the fact that for unknown and frontier words the combined stimuli that provided both definitional and contextual information were easier than were the word pairs that provided only definitional information or only contextual information. Two cues were more effective than one, and the definitional cue was as effective as the contextual cue.

196


Table 2 Mean Proportion Correct and Standard Errors (SEs) on the Word Pair Decision Task in Experiment 5

Method Subjects

Word level Word-pair type

Unknown

Frontier

Known

Definitional Proportion correct

.61

.65

.96

.054

.030

.011

.59

.65

.92

.051

.031

.017

.70

.73

.95

.042

.045

.014

SE Contextual Proportion correct

SE Combination Proportion correct

SE

Analysis of the means indicated that subjects always performed above chance. Even on their unknown words, they were able to choose reliably the correct item, regardless of the type of information provided by that item. This result rules out the contention that subjects would need the complete sentence context to discriminate between the two uses of the word. Apparently, the information necessary to make the choices required here is available from unknown words. Choosing the phrase that satisfied selectional restrictions was no easier than choosing the word that shared semantic attributes with the target. Perhaps these two types of information are acquired together, and the distinction between these types is not as apparent psychologically as formal analyses imply. The current results suggest that people have very abstract information about unknown words. Eysenck's (1979) subjects also seemed to have this level of information, allowing them to perform well on the semantic-differential judgments. Such abstract information allows subjects to perform above chance on either the meaning or the selectional-restriction pairs. For example, knowing or inferring that dowager is animate leads to the correct choice in either condition. This abstract information may be due to classification of the word into a semantic field or neighborhood (Durso & Coggins, 1991; Miller & Gildea, 1987), a process that the current results suggests takes place early in the process of word acquisition. Interestingly, this classification could have taken place during some preexperimental experience with the word, or the classification could have been computed on-line when the subject was forced to make the sentence decision. Experiment 6 attempts to gain some insights into this issue by taking a different methodological tack.

Experiment 6 To gain a better understanding of the types of information available to unknown words and frontier words, we attempted to gain more direct access to the information by asking subjects to free associate to the words on the precheck. We were also able to use the free-association protocols to gain some insight into the source of the information, whether the attribute produced was retrieved from a preexperimental memory or was computed at the time of retrieval.

Subjects were 30 volunteers (Shipley-Hartford = 29) from the same pool. All were native English speakers, and none had participated in previous experiments.

Materials and Procedure Materials were identical to the constraint-only sentence condition used in Experiment 1, with the exception that the 75 items from the precheck were typed individually on 3 x 5 in. index cards for use during the free association task. Subjects were tested individually. After the precheck, subjects performed the free association task. They were presented with the 75 items on index cards one at a time. Subjects were told to say the first thing that came into their minds for each item. They were told that anything was an acceptable response: "If I were to say the word dog you may respond hog because it sounds like dog, or dig because it looks like dog. You may say words that describe dog like bark or lovable or a classification of dog, like animal; or dog may remind you of cat, or urinate on trees because that is what dogs do." After the experimenter had pronounced an item and handed the card to the subjects, the subjects began to give free associations. If the subjects made no response, they were prompted by the experimenter. In addition, if subjects stopped responding, they were prompted by the experimenter saying, "Anything else?" If the subjects made no response to the item within 30 s, the experimenter prompted one final tim^ and then proceeded on to the next card. Subjects were presented with the items in random order, and all responses were recorded on microcassette. Following the free association task, subjects completed the sentence-decision task using the constraint-only sentences.

Results Tapes from 2 subjects were uninterpretable, thus analyses were based on 28 subjects. The initial analysis confirmed that performance was affected by word level, F(2, 26) = 100.47, and that performance was above chance at each level. The level of performance on the sentence decision task fell within the range we have been observing throughout for unknown (.61; SEM = .039), frontier (.73; SEM = .044), and known (.99; SEM = .006) words, and thus the free association procedure did not affect the sentence decision task in any apparent way.

Free Associations Inspection of the free associations for unknown and frontier words revealed responses ranging from simple responses based on the sound or appearance of the target word, such as "orangutan" to the word harangue, to more meaningful responses. Some of the responses that shared some meaning with the target word could have been derived from the physical features of the target: "dowry" to dowager, for example. Other meaningful responses, however, did not derive from the target stimulus in any discernible way. Consider the following protocol for the word vacillate. "Vaseline, I think of leaving. I don't know, like you're a little edgy, kind of

PARTIAL KNOWLEDGE indecisive—you vacillate about something. I don't know, like somebody who is always late." Despite the subject's repeated denials of knowledge of the word, and amidst a number of evidently irrelevant associations, there seems to exist the kernel of some understanding of the word. Subsequent analyses involved classifying the free association responses of subjects for each word prior to determining accuracy on the sentence decision task. Two classifications were attempted. First, we divided the protocols into those that could have been produced because of the physical features of the target and those that could not have been. Second, we divided the protocols into those in which a sophisticated judge could find meaning and those in which the scorer could find no relevant meaning.

Data-Driven Versus Conceptually Driven Responses Data-driven protocols were those in which all relevant responses could be traced to the physical characteristics of the target. Conceptually driven protocols were those in which the subject responded with at least some relevant information that could not be traced to the physical characteristics of the target. These, by implication, were responses based on prior experiences, perhaps memories, of the target. For example, if a subject said "dowry" and "elderly woman" to dowager, the protocol would be classified as conceptually driven because elderly woman could not be derived from the physical target. As another example, if the subject said "vacation" and nothing else to dowager, the response would be classified as conceptually driven, because vacation could not be derived from the sound or look of dowager. We were very liberal in assigning a protocol to the datadriven category. For example, a response of "fancy dress" to the target eloquent was classified as data-driven because the response seemed clearly mediated by the sound-alike elegant. This liberal scoring procedure left in the conceptually driven classification those protocols that could in no way be traced to the physical features of the target. Thus, although there may be more protocols based on memories than this analysis would reveal (for example, a person might have an implicit memory of "dowry" for dowager), we wanted to determine whether any information about unknown words was present that could be attributable to memories and, further, whether performance on these target words would exceed chance. Accuracy could be above chance for either classification. Conceptually driven protocols presumably reflect memories that could lead to above-chance performance on the sentence decision task, but that did not influence classification on the precheck. Thus, these memories may not be ones to which subjects have conscious access, or if they are available to consciousness, the subjects may be unwilling to make the metacognitive assumption that the information is relevant to the familiarity or lexical decisions required Jon the precheck. Data-driven protocols presumably due to iound-alike-lookalike relations to the target could also yield above-chance performance. Morphology and phonology, to the extent they elicit meaning, could be used (see also Koriat & Lieblich, 1974). One of us recalls from graduate school an illustration

197

Figure 1. Sketches of hypothetical tribe members from an example attributed to Wittgenstein. (One tribe is the Kee Kee Kee and one the Marmalo.) that was attributed to Wittgenstein. The illustration is of apocryphal origin, but we continue the attribution, although we have searched in vain for confirmation. The illustration provides a forceful demonstration of how meaningful decisions could be based on the sound or look of a word. Wittgenstein's example asks that one pretend to be an anthropologist who has discovered the two tribes illustrated in Figure 1. One of the tribes is called the Kee Kee Kee and one the Marmalo. People are virtually unanimous in assigning the label Kee Kee Kee to the hard-edged tribesman on the left. The power of Wittgenstein's example is the general agreement achieved when people are asked to match labels they have never heard with pictures they have never seen. Because language is a system, subjects might take advantage of this system and attempt to make meaningful decisions based on the physical stimulus. Of course, the assignment of a label to a referent is arbitrary for any particular assignment (de Saussure, 1959), but as the system develops in adults, correlations in the system—sometimes accurate, sometimes misleading— will be available to the system user. This contention is a weaker version of what was known as phonetic symbolism (e.g., Brown, 1958;Sapir, 1929). Of the total number of protocols for unknown words, 16% were conceptually driven, 74% could be classified as datadriven, and 10% did not yield any response. For frontier words, 32% were conceptually driven, 63% could be classified as data-driven, and 5% were no-responses. Mean accuracy for each subject was computed. Mean accuracy and the number of subjects contributing to each mean appear in Table 3. In all four cases, performance was above chance. Although accuracy appears somewhat higher for targets that yielded conceptually driven protocols, analyses' did

1 Pairwise comparisons of conditions in which ns differed, but in which the same subject could have contributed to both conditions, were analyzed in three ways: A between-subjects comparison ignored the covariance; a correlated test dropped subjects who failed to contribute to both conditions; and a comparison that used all subjects to estimate means, but only the repeated subjects to estimate covariance and degrees of freedom. All tests led to identical conclusions.

198


Table 3 Mean Proportion Correct and Standard Errors (SEs) on the Sentence Decision Task for n Subjects Contributing to a Classification of the Free-Association Protocols Word level Unknown Protocol type Frontier Conceptually driven' Proportion correct .68 .77 SE .081 .066 17 20 n Data driven6 .71 .59 Proportion correct .046 .057 SE 27 n 28 * Those protocols that could not have been due to the sound or look of the target word. b Those protocols that could have been due to the sound or look of the target word.

not confirm this. Thus, subjects performed above chance for targets that brought only sound-alike-look-alike associations to mind and for targets that brought other associations, presumably memories, to mind.

Meaningful Versus Nonmeaningful Responses The same protocols were also classified into meaningful and nonmeaningful responses. Judges blind to the responses of the subject, but with access to the definitions of the words, attempted to find meaning in the subjects' protocols. Meaningful responses included those closely related to the meaning of the target, broadly associated to the meaning (e.g., "marriage" to dowager) and antonyms to the meaning (e.g., "newlywed" to dowager). Nonmeaningful responses included contextual responses (e.g., "queen" to dowager), sound-alikes, and idiosyncratic responses. Of the total number of protocols for unknown words, the scorers could find meaning in 15% of the subject responses, and for frontier words, in 31% of their responses. Mean accuracy was computed for each subject. Means for each classification and the number of subjects contributing to those means appear in Table 4. Performance was above chance, unless the scorers could not find meaning for the unknown words. If the sophisticated observer could find meaning in the protocol of an unknown word, subjects performed significantly better than if the scorer could not find meaning. Apparently, subjects have information about the meaning of unknown words. As with the TOT effect, information is activated in memory that could be used to help retrieve the target word, but subjects do not have the metamemorial sophistication to use the information. If this speculation is correct, then subjects should, on occasion, choose to omit certain items on the sentence decision task if given the opportunity. Yet, if they were subsequently forced to respond to those omitted items, they would again perform above chance. Experiment 7 tests exactly this speculation.

Experiment 7 Method Subjects Subjects were tested until data from 30 native English speakers from the introductory psychology subject pool could be used. We replaced 3 subjects who did not have words at each level of word knowledge. None had served as subjects in previous experiments. Shipley-Hartford scores averaged 30.

Materials The materials included the precheck, the distractor task, and the constraint-only sentences. For the sentence-decision task, each sentence pair was typed on a 5 x 7 in. card. On half of the cards, the correct sentence appeared above the incorrect sentence, and on the other half of the cards, the order was reversed.

Procedure Subjects completed the Shipley-Hartford, the precheck, and the distractor task. They then went on to the sentence-decision task. For this task, subjects were presented with each sentence pair individually. They were told to choose the sentence that used the underlined word correctly. In addition, they were told that for each correct choice they would receive 10*. If they chose the incorrect sentence, they would lose 10*. They were also given the option to pass on any sentence pair if they were not confident enough of their choice to risk the dime. They were explicitly told that skipping a sentence pair would neither cost nor yield them any money. The monetary contingencies were incorporated into this experiment so that there would be real consequences associated with attempting or omitting an item. No subjects were aware that they would be required later to make a choice on the sentence pairs they skipped. All subjects were given 50* to begin. When subjects chose the correct sentence, they were immediately given a dime by the experimenter. If the incorrect sentence was chosen, a dime was immediately taken away from the subjects. Any sentence pairs the subjects decided to skip were set aside by the experimenter. After subjects had made their first pass through the sentence pairs, they were told that the "money round" was over. Any money they had earned was theirs to keep. The experimenter then shuffled the

Table 4 Mean Proportion Correct and Standard Errors (SEs) on the Sentence Decision Task for n Subjects Contributing to a Classification of the Free Association Protocols Word level Frontier Protocol type Unknown Meaning" Proportion correct .82 .78 SE .067 .069 17 22 n No meaning .71 Proportion correct .57 SE .047 .055 n 28 28 " Those protocols for which knowledgeable scorers could find meaning relevant to the word.

PARTIAL KNOWLEDGE cards containing the skipped sentence pairs and told the subjects that they must now decide which sentence of each remaining pair used the target word correctly. These remaining cards were then presented individually, and subjects indicated their choice on each.

Results Mean proportions correct on the sentence decision task were computed as before. The first analysis was a one-way MANOVA collapsed across pass. This analysis was done to determine whether the main effect of word level would replicate using this methodology, which it did, F(2, 28) = 39.01. In addition, means were significantly above chance for known (.95; SEM = .015), frontier (.73; SEM = .029), and unknown (.71; SEM = .040) words, again replicating previous results found with the constraint-only sentence pairs, although as in Experiment 5 frontier words and unknown words were comparable. Of more interest was the performance on those items the subjects did not attempt when they were given a choice. Performance was above chance on the second pass for both unknown and frontier words (see Table 5). Thus, although subjects chose to omit these items during the initial pass through the stimuli, their ultimate performance on the items was better than would be expected by guessing alone. In fact, there is some evidence in Table 5 that subjects could perform as well on the second pass as on the first. This suggests that subjects had difficulty in deciding on which items to bet. Apparently, the subjects were unable, or unwilling, to use the information they retrieved to make a decision about the sentence pair. This result is consistent with the speculations from Experiment 6 in which it seemed that subjects had relevant information, but were unable to use the information effectively, either to pick between sentences or to affirm the lexicality of the word on the precheck.

General Discussion Across a number of experiments, we observed a surprising level of performance (cf. Eysenck, 1979; Koriat & Lieblich, 1974; Loewenthal, 1971; Shore & Durso, 1990) on items our subjects claimed were not English words. In fact, this level of

Table 5 Mean Proportion Correct and Standard Errors (SEs) on the Sentence Decision Task for n Subjects Freely Attempting the Item (First Pass) and Those Forced to Attempt the Item (Second Pass) Word level Pass First Proportion correct SE n Second Proportion correct SE n

Unknown

Frontier .70 .046

.75 .052 30

26

.78 .052

.70 .047 24

22

199

performance was sometimes (Experiment 5; Experiment 7, first pass) comparable to performance on words that subjects claimed were vaguely familiar. Words that subjects denied were part of the English language allowed subjects to choose between a correct and anomalous use of the word, between a synonym and an unrelated word, and between a word that satisfied selectional restrictions and one that did not. This level of performance was not due to information contained in the sentence rather than the word: Sentence frames without the word were equally acceptable (Experiment 2) and abovechance levels of performance were observed when testing did not involve sentences at all (Experiment 5). The above-chance performance was not attributable to a mismatch between study and test, either in terms of stimuli or in terms of the task required of the subject: When subjects were given the sentences as context in which to make the lexical decision (Experiment 3), performance on words they continued to deny were not English was again above chance; when subjects were given the opportunity to omit a sentence pair (Experiment 7), they nevertheless performed above chance on those items when the experimenter forced them to choose. Thus, even when they were given the opportunity to perform the sentence decision task and chose not to take it, they ultimately performed well. Successful performance on unknown words, however, was not always observed. Apparently, subjects required a contrast between a correct and an incorrect use; they could not make absolute judgments about the correct use of unknown words (Experiment 4). In addition, the contrast could not be too subtle; subjects performed at chance when forced to reject a foil that differed from the correct sentence in specific details of the word's meaning (Experiment 1). Subjects' ability to perform these tasks for even unknown words appears due to both memories of prior experiences with the word and to information supplied by the language system (Experiment 6). When free association protocols contained nothing that could be considered formally related to the target, we assumed that the protocols reflected something other than data-driven associations. A likely candidate are memories of some type. It is tempting, given current thinking, to assume that these memories are implicit because they apparently were not available to consciousness when subjects made the lexical decision on the precheck. The fact that the subjects will omit items on which they will later perform above chance also could be interpreted to mean that subjects have information about the word of which they are not consciously aware. However, the free association data suggest that this conclusion is premature. Subjects often could articulate information relevant to the meaning of words that they denied were part of their language. In the light of this result, an alternative to implicit memory is that subjects were unwilling to use any information that may have entered consciousness to confer lexical status on the word or to choose between two sentences. In this view, a letter string (whether a word or not) activates information in the lexicon, and subjects must decide whether information is present that suggests familiarity. Failing that, subjects must decide whether information is present that would justify conferring lexicality on the string. If not, the

200


string will be judged as a nonword. If the letter string is, in reality, a real English word, some of the information might be part (albeit a nascent part) of the meaning of the word, either because the word was experienced earlier or because the letter string activates neighbors in a lexical neighborhood in which a correspondence exists between sound and general meaning. For unknown words, the meaningful information is not the type thought to elicit a feeling of familiarity (e.g., Mandler, 1980). In addition, it is likely to be irrelevant to the task of making a lexical decision, given that nonwords can also activate similar sets of semantic information. The findings from Experiment 7, cited earlier as consistent with an implicit memory interpretation, are thus also consistent with the idea that subjects often choose to omit items they will ultimately answer correctly because they feel uncomfortable discriminating relevant from irrelevant information (see Steinberg & Powell, 1983). Thus, although the information available to unknown words does not allow subjects to confer familiarity or even lexicality on the string, it may provide enough semantic attributes to make relatively simple decisions, like goodness or badness (Eysenck, 1979), or, as in the current article, contrasts between a correct use and a very wrong use. These few attributes are not sufficient, however, for subtle contrasts (Experiment 1) or absolute judgments (Experiment 4). In fact, these few attributes are insufficient to persuade subjects to make a voluntary response (Experiment 7), despite the fact the attributes are useful in distinguishing between the two uses of the word in the sentence pairs. Unlike unknown words, performance on frontier words was always above chance. The fact that frontier words, but not unknown words, yielded above-chance performance with little or no contrast suggests that frontier words and unknown words do activate different types of information. Explanations that contend that the only difference between the two types of words is the general feeling of familiarity of the letter string would have difficulty explaining this result. The two levels also differ in the type of knowledge they make available. The difference in type of knowledge appears to be other than the meaning and selectional-restriction information isolated in Experiment 5. Instead, we have been entertaining the possibility that frontier words contact specific episodic contexts more often than do unknown words. This speculation helps explain why frontier words usually surpass unknown words when the retrieval cue is a sentence, but not when the retrieval cue is a word or short phrase. A sentence might contact a specific episodic memory for the subject that a synonym or short phrase would not; frontier words presumably have such episodes, whereas unknown words presumably would not. Bailey and Shore (1990) provide evidence that frontier words contact episodic contexts more often than do unknown words. They classified the free associations of Experiment 6 according to a scheme very similar to that proposed by O'Connor (1940). The only difference in accuracy between frontier and unknown words that they report was for free associations classified as O'Connor's episodic context (e.g., "queen" in response to dowager). To be classified as context, the free association protocol must have been specific, but not

close enough in meaning to be classified as a meaning protocol. For frontier words, subjects may retrieve specific instances, including cognates like queen, or slightly off scenarios in which the word could appear. If Bailey and Shore's (1990) speculation proves to be correct, it would make sense in the context of our discussion. In addition to the information available to unknown words, frontier words contact particular experiences in the subjects' history with the word. These histories are useful when the retrieval cue recapitulates those histories. Thus, the impoverished context of our word pairs would be unlikely to contact the previous episodic contexts in which the word had been experienced. In fact, even when we created the selectional restrictions, we made certain that the modifier bore an arbitrary relationship to the target. We did this to separate meaning and selectional-restriction information more clearly, but as a consequence the stimuli made certain that subjects could not use any particular experience to make the discrimination. On the other hand, sentences are more likely, but are not guaranteed, to contact one of the particular experiences that is part of the frontier word's representation in memory. Thus, we would expect sentences using frontier words typically, but not necessarily, to yield better performance than sentences using unknown words. Regardless of the ultimate validity of some of the details discussed here, we have explored a phenomenon of partial knowledge that has been noted in a number of literatures, but that has received little empirical attention. Exploration of partial knowledge can have direct consequences, not only for instruction, but for extant models of, what researchers assume is, complete knowledge. Both structural and metamemorial influences are important for a complete understanding of partial knowledge.

References Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological Review, 89, 369-406. Bailey, K. A., & Shore, W. J. (1990, April). Implicit knowledge: Categorizing free associations of unknown words. Paper presented at the 36th Annual Meeting of the Southwestern Psychological Association, Dallas, TX. Blake, M. (1973). Prediction of recognition when recall fails: Exploring the feeling-of-knowing phenomenon. Journal of Verbal Learning and Verbal Behavior, 12, 311-319. Brown, R. (1958). Words and things. New York: Free Press. Brown, R., & McNeill, D. (1966). The "tip-of-the-tongue" phenomenon. Journal of Verbal Learning and Verbal Behavior, 5, 325337. Chase, W. G., & Simon, H. A. (1973). The mind's eye in chess. In W. G. Chase (Ed.), Visual information processing (pp. 215-281). San Diego, CA: Academic Press. Clark, E. V. (1973). What's in a word? On the child's acquisition of semantics in his first language. In T. E. Moore (Ed.), Cognitive development and the acquisition of language, (pp. 65-110). San Diego, CA: Academic Press. Curtis, M. E. (1987). Vocabulary testing and vocabulary instruction. In M. G. McKeown & M. E. Curtis (Eds.), The nature of vocabulary acquisition (pp. 37-51). Hillsdale, NJ: Erlbaum. Dale, E. (1965). Vocabulary measurement: Techniques and major findings. Elementary English, 42, 395-401. de Saussure, F. (1959). Course in general linguistics. New York:

PARTIAL KNOWLEDGE McGraw-Hill. Drum, P. A. (1983). Vocabulary knowledge. In J. A. Niles & L. A. Harris (Eds.), Searches for meaning in reading/language processes and instruction (pp. 163-171). Rochester, NY: National Reading Conference. Drum, P. A., & Konopak, B. C. (1987). Learning word meanings from written context. In M. G. McKeown & M. E. Curtis (Eds.), The nature of vocabulary acquisition (pp. 73-87). Hillsdale, NJ: Erlbaum. Durso, F. T., & Coggins, K. A. (1991). Organized instruction for the improvement of word knowledge skills. Journal of Educational Psychology, 83, 108-112. Eysenck, M. W. (1979). The feeling of knowing a word's meaning. British Journal of Psychology, 70, 243-251. Hart, J. T. (1965). Memory and the feeling-of-knowing experience. Journal of Educational Psychology, 56, 208-216. Hart, J. T. (1967). Memory and the memory-monitoring process. Journal of Verbal Learning and Verbal Behavior, 6, 685-691. Hayes-Roth, B. (1977). Evolution of cognitive structures and processes. Psychological Review, 84, 260-278. Hintzman, D. L. (1986). "Schema abstraction" in a multiple-trace memory model. Psychological Review, 93, 411-428. James, W. (1890). The principles of psychology (Vol. 1). New York: Dover. Jenkins, J. R., Stein, M. L., & Wysocki, K. (1984). Learning vocabulary through reading. American Educational Research Journal, 21, 767-787. Koriat, A., & Lieblich, I. (1974). What does a person in a "TOT" state know that a person in a "don't know" state doesn't know? Memory & Cognition, 2, 647-655. Loewenthal, K. (1971). A study of imperfectly acquired vocabulary. British Journal of Psychology, 62, 225-233. Lord, F. M., & Novick, M. R. (1968). Statistical theories of mental test scores. Reading, MA: Addison-Wesley. Lovelace, E. (1987). Attributes that come to mind in the TOT state. Bulletin of the Psychonomic Society, 25, 370-372. Mandler, G. (1980). Recognizing: The judgment of previous occurrence. Psychological Review, 87, 252-271. Miller, G. A. (1978). Semantic relations among words. In M. Halle, J. Bresnan, & G. A. Miller (Eds.), Linguistic theory and psychological reality (pp. 60-118). Cambridge, MA: MIT Press. Miller, G. A., & Gildea, P. M. (1987). How children learn words. Scientific American, 257, 94-99. Nagy, W. E., Herman, P. A., & Anderson, R. C. (1985). Learning words from context. Reading Research Quarterly, 20, 233-253. Nelson, K. (1982). The syntagmatics and paradigmatics of conceptual development. In S. Kuczaj II (Ed.), Language development (Vol 2., pp. 335-364). Hillsdale, NJ: Erlbaum.

201

Nelson, T. O., Fehling, M. R., & Moore-Glascock, J. (1979). The nature of semantic savings for items forgotten from long-term memory. Journal of Experimental Psychology: General, 108, 225250. Nelson, T. O., Leonesio, R. J., & Landwehr, R. S. (1986). A comparison of three predictors of an individual's memory performance: The individual's feeling of knowing versus the normative feeling of knowing versus base-rate item difficulty. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 279-287. Nelson, T. O., Leonesio, R. J., Shimamura, A. P., Landwehr, R. F., & Narens, L. (1982). Overlearning and the feeling of knowing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 8, 279-288. Nelson, T. O., & Narens, L. (1980). Norms of 300 general-information questions: Accuracy of recall, latency of recall, and feeling-ofknowing ratings. Journal of Verbal Learning and Verbal Behavior, 19, 338-368. O'Connor, J. (1940). Unsolved business problems. Boston, MA: Johnson O'Connor Research Foundation. Sapir, E. (1929). A study in phonetic symbolism. Journal of Experimental Psychology, 12, 225-239. Schacter, D. L. (1983). Feeling of knowing in episodic memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 9, 39-54. Shore, W. J., & Durso, F. T. (1990). Partial knowledge in vocabulary acquisition: General constraints and specific detail. Journal of Educational Psychology, 82, 315-318. Sines, L. K. (1958). Intelligence test correlates of Shipley-Hartford performance. Journal of Clinical Psychology, 14, 399-404. Sternberg, R. J., & Powell, J. S. (1983). Comprehending verbal comprehension. American Psychologist, 38, 878-893. Trembly, D. (1966). Laws of learning general and specialized vocabularies [Summary]. Proceedings of the 74th Annual Convention of the American Psychological Association, 1, 229-230. Tulving, E. (1979). Relation between encoding specificity and levels of processing. In L. S. Cermak & F. I. M. Craik (Eds.), Levels of processing in human memory, (pp. 405-428). Hillsdale, NJ: Erlbaum. Tulving, E., & Thompson, D. M. (1973). Encoding specificity and retrieval processes in episodic memory. Psychological Review, 80, 352-373. Wilkinson, A. C. (1982). Theoretical and methodological analysis of partial knowledge. Developmental Review, 2, 274-304. Winer, B. J. (1971). Statistical principles in experimental design (2nd ed.). New York: McGraw-Hill. Yarmey, A. D. (1973). I recognize your face but I can't remember your name: Further evidence on the tip-of-the-tongue phenomenon. Memory & Cognition, 1, 287-290.

(Appendix follows on next page)

202


Appendix A List of Words and Nonwords Used in All Experiments Words aesthetic ambiance ambivalent anathema ancillary aplomb appropriate armada benefactor cachinnate catalyst cincture coherent consummate contiguous cynical demeanor depict

diffident disconcerted dowager dubious eloquent emanate embellish ensconce eulogy expedite facet gratuitous harangue horrendous imminent lucid ludicrous maudlin melancholy obsequious

obtuse paucity pertinent pervasive placate pretext quintessential redundant refulgent rejuvenate relinquish reprimand scintillate skeptic sloth stipulation sublime susceptible umbrage uxorious

vacillate vilify

Nonwords calsar clowex devernal drallen edarthic fossern garseth hiliathum jandelar merriton phisteron redistac teightran tessach thonstan

Appendix B Examples of Materials Correct and Incorrect Sentences Used in the Constraint-Only Pairs He thought the joke was so funny that he couldn't help but cachinnate. The cake began to cachinnate in the oven. The dowager has her dead husband to thank for her great wealth. The circus performer could juggle the dowager and two bowling pins. The minister gave a stirring eulogy at the funeral. The eulogy got up and walked out of the office. The obsequious man would not quit flattering me. The man sneezed through the obsequious kleenex. The sun was refulgent off the water. The man's bad breath was refulgent. Comedians constantly vilify the president's ability as a leader. The settlers began to vilify as soon as they reached the river.

Incorrect Sentences Used (Together With Correct Sentences Above) to Form the Specific, Detail Pairs (Experiment 1) The widow began to cachinnate during the funeral. The dowager and her husband recently celebrated their fiftieth wedding anniversary. The eulogy given by his friend made everyone think less of the man. The obsequious group started a riot when they failed to follow the anarchist's orders.

The murky pond had a refulgent quality. The professor used a compliment to vilify his assistant.

Definitional, Contextual, and Combined Pairs (Experiment 5) laugh-rise audiences cachinnate-cak.es cachinnate audiences laugh-cakes rise widow-plate friendly dowager-broken dowager friendly widow-broken plate tribute-client stirring eulogy-busy eulogy stirring tribute-busy client compliant-cheap obsequious colleague-obsequious home compliant colleague-cheap home offensive-gleaming refulgent sun-refulgent breath gleaming sun-offensive breath slander-squeak comedians vilify-chairs vilify comedians slander-chairs squeak

Received August 8, 1990 Revision received November 16, 1990 Accepted December 4, 1990