Neighborhood Density, the Tip-of-the-Tongue Phenomenon, and Aging

A tip-of-the-tongue (TOT) elicitation task was used with younger adults in Experiment 1 and with older adults in Experiment 2 to examine the influence of word frequency, neighborhood density, and neighborhood frequency on the retrieval of phonological word forms from the lexicon. The results of Experiment 1 replicated the results of Harley and Bown (1998): More TOT states were elicited for words with low frequency and sparse neighborhood density (i.e., words with few similar sounding words). However, the results from Experiment 2 showed that in addition to word frequency and neighborhood density, neighborhood frequency, the mean frequency of phonological neighbors, also influenced lexical retrieval for speech production in older adults. Specifically, neighborhood-frequency interacted with word frequency and with neighborhood density. More TOT states were elicited for words with low neighborhood frequency and low word frequency. In addition, more TOT states were elicited for words with low neighborhood-frequency and sparse neighborhoods. These results demonstrate that the number of similar sounding words affects lexical retrieval in production as well as perception. Furthermore, the influence of these lexical characteristics on the process of retrieving word-forms during speech production changes with age. The results are discussed within the context of the Node Structure Theory (MacKay, 1987), a model of cognitive processing. Introduction Speech production is a rapid and efficient process. However, there are instances in which the fluent retrieval of a lexical item fails to occur. One example of failed retrieval occurs in tip-of-the-tongue (TOT) states. Tip-of-the-tongue states occur when only partial information associated with a word can be retrieved: information regarding the meaning or syntactic class of the word may be accessible, but not the complete phonological form of the word. The ability to access partial information often results in one having a “feeling of knowing” the word, despite being unable to retrieve all the information associated with the word. Factors that Affect Lexicalization and TOTs The process of retrieving a word form from lexical memory during speech production is called lexicalization. Several factors affect the speed and accuracy of lexicalization. The factors that affect the speed and accuracy of normal, unimpaired, lexicalization also affect instances of incomplete lexicalization, or TOT states. One factor that affects the speed and accuracy of normal lexicalization, as well as TOT states, is word frequency. Using a picture-naming task, Oldfield and Wingfield (1965) demonstrated that pictures of high-frequency words were named more quickly than were pictures of low-frequency words. Stemberger (1985; Stemberger & MacWhinney, 1986; see also Dell, 1990) found that phonological speech errors, such as spoonerisms (switching the initial phoneme of two nearby words, such as “darn bore” instead of “barn door”), perseverations (carrying-over a phoneme from a word produced earlier, such as “Ladies and lentlemen...” instead of “Ladies and gentlemen...”), and anticipations (producing a phoneme 3 This is about the only point that researchers agree on. Debates continue about the processes involved in speech production. For example, Garrett (1976) and Levelt (1989) take a modular approach to processing, whereas Dell (1986) and Harley (1984; 1993) argue for an interactive approach. There are also debates about where certain types of information are represented in the speech production system. For example, Roelofs, Meyer, and Levelt (1998) argue that syntactic information is accessed at the lemma level, whereas Caramazza and Miozzo (1998) claim that syntactic information is accessed at the lexeme level. NEIGHBORHOOD DENSITY AND TOTS 133 from an upcoming word, such as “pig pig” instead of “big pig”), occurred more often for low-frequency words than for high-frequency words. Finally, more TOT states occur and are experimentally induced in low frequency words than in high frequency words (Brown & McNeill, 1966; Burke, MacKay, Worthley, & Wade, 1991; Harley & Bown, 1998; cf. Yaniv & Meyer, 1987). Taken together, these findings suggest that the well-documented perceptual disadvantage of low-frequency words (i.e., poorer identification and slower response times) is paralleled by a similar disadvantage in production, as evidenced by greater difficulty with lexicalization. Another factor that affects the speed and accuracy of both perception and lexicalization is neighborhood density. Neighborhood density refers to the number of words that are phonologically similar to a given target word (Luce & Pisoni, 1998). A rough measure of phonological similarity can be obtained by determining the number of new words that are created by the addition, deletion, or substitution of a phoneme in a target word. For example, the word “cat” has as neighbors the words “scat,” “at,” “hat,” “cut,” and “cap,” as well as other words. Words with many similar sounding words are said to have dense neighborhoods, whereas words with few similar sounding words are said to have sparse neighborhoods. Previous research has shown that words with sparse neighborhoods are recognized more quickly and more accurately than words with dense neighborhoods (Luce & Pisoni, 1998). Previous investigations (Goldinger & Summers, 1989; Harley & Bown, 1998; Vitevitch, 1997a; 1997b) have also demonstrated that neighborhood density can affect speech production. For example, Goldinger and Summers (1989) found that neighborhood density influenced the voice onset time (VOT) for spoken words. VOT refers to the point in time at which vocal fold vibration starts, following the release of a closure (Crystal, 1992). Participants repeated word pairs that differed in the voicing of the initial consonants within the pair (e.g., dutch-touch) and that varied in neighborhood density across pairs. An acoustic analysis showed that the differences in VOT between the first word and the second word of the pairs were larger for word pairs with dense neighborhoods than for word pairs with sparse neighborhoods. These differences decreased across sessions for word pairs with sparse neighborhoods, but increased across sessions for word pairs with dense neighborhoods. Furthermore, Goldinger and Summers found that the interword interval, or the time between the offset of the first word and the onset of the second word within each minimal pair, varied with neighborhood density. The interword interval was greater for dense neighborhood word pairs than for sparse neighborhood word pairs. These results demonstrate that neighborhood density influences certain aspects of timing in speech production. The accuracy of lexical retrieval in speech production is also affected by neighborhood density. Vitevitch (1997a) used tongue twisters containing words that had either dense or sparse neighborhoods to elicit phonological speech errors from participants. He found that more errors occurred in tongue twisters that contained words with sparse neighborhoods than with dense neighborhoods. These results suggest that neighborhood density influences speech production in demonstrable ways (see also Vitevitch, 1997b; Wright, 1997; cf. Jescheniak & Levelt; 1994). Specifically, neighborhood density appears to produce “supportive” or facilitative effects among words. That is, words with many similar sounding words (a dense neighborhood) are produced more accurately than words with few similar sounding words (a sparse neighborhood). These findings contrast with the competitive effect of neighborhood density typically observed in spoken word recognition: Words with dense neighborhoods are recognized more slowly and less accurately than words with sparse neighborhoods (Luce & Pisoni, 1998). 4 An alternate method used to measure "similarity" is to use phoneme confusion matrices as in Luce and Pisoni (1998) in the calculation of Neighborhood Probability Rules (NPRs). Both methods have been successfully used to demonstrate effects of neighborhood density on spoken word recognition. VITEVITCH AND SOMMERS 134 Vitevitch (1997a; see also MacKay & Burke, 1990) speculated that the difference in neighborhood density effects found in speech production and spoken word recognition were due to the differences in the flow of information during speech production and spoken word recognition. That is, in spoken word recognition, acoustic-phonetic input activates many similar sounding words in memory (e.g., Luce & Pisoni, 1998). This candidate set must be winnowed down to a single item that will then retrieve semantic and syntactic information related to that word from the lexicon. Thus, more time will be required to winnow down the candidate set if there are many competitors (Luce & Pisoni, 1998). In contrast, speech production begins with a single conceptual representation that proceeds to activate a single lexical item and a single phonological word form (Levelt, 1989). That single phonological word form then activates the many sub-lexical units that it contains, such as syllables, phonemes, features, etc. Thus, a word that has components shared by many other words (a word with a dense neighborhood) will be able to spread activation along pathways between those components that are well traversed. A word that has components shared by few other words (a word with a sparse neighborhood) will have difficulty spreading activation along the pathways between components that are less traveled. Additional evidence of neighborhood density affecting speech production can be found in the work of Harley and Bown (1998). They recently reported that more TOT states were elicited for words from sparse neighborhoods than for words from dense neighborhoods, and also suggested that neighborhood density played a “supportive” role in speech production (Harley & Bown, 1998). Although Harley and Bown (1998) accounted for their results in the context of extant interactive models of lexicalization (i.e., Dell, 1986; Harley, 1993), their results are difficult to clearly interpret because of confounding variables in their stimulus set. Specifically, in two experiments that manipulated word frequency and neighborhood density, Harley and Bown (1998) attempted to induce TOT states experimentally using words that varied in length from one syllable (e.g., “act”) to five syllables (e.g., “chronological”). Word length was a variable that was not stringently controlled in their stimuli, and, unfortunately, proved to be a confounding variable. The results of their first experiment showed that more TOT states were reported for words that were low in frequency and that had few neighbors as defined by Coltheart-N (Coltheart, Davelaar, Jonasson, & Besner, 1977). Although the tip-of-the-tongue phenomenon is often described as an inability to retrieve a soundbased representation from the lexicon, Harley and Bown (1998) constructed their stimulus set using a metric of similarity based on orthographic similarity (Coltheart-N) instead of a metric based on phonological similarity. It should be noted, however, that when Harley and Bown analyzed the results from a reduced set of their stimuli based solely on phonological neighborhoods, their findings remained relatively unchanged. However, when Harley and Bown performed a regression analysis on the data in Experiment 1, they found a significant effect of word length on TOT states: TOT states were more likely to occur with longer words than shorter words. Across the lexicon, short words tend to have denser neighborhoods than longer words (Bard & Shillcock, 1993; Pisoni, Nusbaum, Luce & Slowiaczek, 1985). Their results are further complicated by other relationships among word frequency, word length, and neighborhood density in the lexicon. For example, Zipf (1965) observed an inverse relationship between word length and word frequency in English: Short words are more common in English than long words. Also, Landauer and Streeter (1973) found a positive correlation between word frequency and neighborhood density: High frequency words tend to have denser phonological neighborhoods than low frequency words. Thus, it is unclear whether the results in Experiment 1 of Harley and Bown (1998) were due to neighborhood density or another related variable. NEIGHBORHOOD DENSITY AND TOTS 135 Harley and Bown (1998) attempted to control word length more precisely in their second experiment by using monosyllabic and disyllabic words (however, the trisyllabic word “audience” appears as a stimulus item in a low N condition) to examine the effects of word frequency and neighborhood density on TOT states. Although the word frequency and neighborhood density effects from Experiment 1 were replicated, a close examination of the stimuli in Experiment 2 reveals that word length was not entirely controlled. An analysis of the stimuli in appendix B of Harley and Bown (1998) shows that words with dense neighborhoods were still shorter than words with sparse neighborhoods. This is true when word length is measured in number of phonemes (dense words, mean = 3.17 phonemes; sparse words, mean = 5.07 phonemes; F (1,58) = 54.15, p < .001) and in number of syllables (dense words, mean = 1.06 syllables; sparse words, mean = 1.83 syllables; F (1,58) = 8.82, p < .001). Given the complex relationships among word length, word frequency, and neighborhood density, it is unclear how each of these individual factors affected TOTs in Harley and Bown (1998). Accounts of TOTs Several hypotheses have been advanced to account for the occurrence of TOT states. One hypothesis states that similar sounding words interfere with— or “block”—the retrieval of the phonological word-form (Jones, 1989, Jones & Langford, 1987; Maylor, 1990; Woodworth, 1929). For example, Jones (1989) presented definitions to participants and primed them with a word that was semantically, phonologically, or both semantically and phonologically related to the target word. Jones (1989; see also Jones & Langford, 1987, and Maylor, 1990) found that more TOT states were elicited when a phonologically related prime was presented after hearing the definition of the target word. Jones (1989) interpreted these results as being consistent with the hypothesis that phonologically related words block the retrieval of the desired word-form. An alternative explanation of TOTs claims that insufficient activation results in incomplete retrieval of the target word (Brown, 1991; Burke, MacKay, Worthley & Wade, 1991). According to this hypothesis, similar sounding words should act to aid rather than block the retrieval of word-forms. Evidence for this hypothesis comes from the work of Meyer and Bock (1992) and Perfect and Hanley (1992). Meyer and Bock (1992) and Perfect and Hanley (1992) showed that the targets used by Jones (1989) differed across conditions in the susceptibility to TOT states. When targets with equal susceptibility to TOT states were used across conditions, phonological primes did not interfere with the retrieval of the target word form; rather, phonological primes aided in the retrieval of the target word-form (Meyer & Bock, 1992; Perfect & Hanley, 1992). The results of Harley and Bown (1998) also support the hypothesis that phonological similarity can serve to support lexicalization. They found more TOTs for words with sparse neighborhoods than for words with dense neighborhoods, suggesting that the more neighbors a word has, the more “support” it receives, and the more likely it will be correctly and completely retrieved. Harley and Bown (1998) accounted for their results by hypothesizing that the representation of an intended word was not fully activated because of insufficient amounts of supportive feedback between the lexeme level (which contains phonological information) and the lemma level (which contains semantic and/or syntactic information). They state that “...[l]emmas corresponding to phonological forms that have no or few close neighbours can receive little or no supporting activation from feedback between the phonological and lemma levels from 5 As in Experiment 1, Harley and Bown (1998) performed a regression analysis, but failed to find a relationship between word length and number of TOT states. However, the restricted range of word length in Experiment 2 (mostly monoand disyllabic words, with one trisyllabic word) compared to the broader range of word length in experiment 1 (words with one to five syllables) may account for the non-significant regression. VITEVITCH AND SOMMERS 136 related forms.” (Harley & Bown, 1998; pp. 163-164). Harley and Bown (1998; see also Harley & MacAndrew, 1992) further hypothesize that weak representations or random noise in the connections between the lemma and phonological representations may also contribute to TOT states. Rather than being at the interface between the lemma and lexeme, as postulated by Harley and Bown (1998), the locus of TOT states may instead be at the interface between the lexeme (i.e., the phonological representation of the whole word) and sub-lexical representations. This hypothesis was postulated by Burke et al. (1991) within the context of the Node Structure Theory (NST), an interactive model of cognitive processing (see MacKay, 1987). Specifically, they state that “...TOTs result when phonological feature nodes receive insufficient priming to become activated.” (Burke et al., 1991, pp. 547). Insufficient activation between “word” and “phoneme” representations as postulated in NST can also explain the results of Harley and Bown (1998). Node Structure Theory and TOTs NST consists of a network of processing units, or nodes, organized hierarchically into semantic, phonologic, and motoric levels. Nodes are localist representations and are connected symmetrically —both bottom-up and top-down connections. The same network of nodes is involved in the perception and production of language (MacKay, 1987). Two processes operate in NST: priming and activation. Priming is the sub-threshold excitation of a node that prepares it for activation. Activation is an all-or-none state in which the node has crossed a certain threshold. Priming has several characteristics. It spreads in parallel to all connected nodes higher and lower in the hierarchical structure of nodes. Nodes can sum the priming that they receive simultaneously from several other nodes, or that they receive temporally across a single connection. Finally, transmission of priming becomes less efficient when a node has been satiated after prolonged and repeated activation. In NST, activation is different from priming. For example, activation does not “spread” as it does in other network theories (e.g., McClelland & Rumelhart, 1981). Instead, activation proceeds sequentially and hierarchically through the network in a top-down and left-to-right manner. Activation must occur (i.e., the threshold must be crossed) in order to consciously retrieve the information associated with a node. In addition to the nodes representing information at various levels, there are sequence nodes that connect nodes that share the same syntactic function or sequential privilege of occurrence in words and sentences. (This collection of similar nodes is referred to as a domain.) When a sequence node is activated, it multiplies the priming of all the nodes in a domain. The consequence of this multiplication of priming is that the node that initially had the most priming in that domain reaches threshold first and becomes activated. This domain specific activation accounts for the regularity found in many types of substitution errors. For example, nouns often substitute for nouns rather than verbs in a sentence, initial consonants often substitute for initial consonants rather than vowels or final consonants, etc. (e.g., Dell, 1986; Stemberger, 1985; MacKay, 1979). Burke et al. (1991) suggest that TOT states arise in NST due to a deficit in transmission of priming across certain connections that are crucial for producing a target word. In a TOT state, semantic nodes become activated giving access to semantic information associated with that word. However, priming NEIGHBORHOOD DENSITY AND TOTS 137 does not spread sufficiently among phonological nodes, resulting in some phonological information not being activated and retrieved. Transmission deficits may be caused by three factors: frequency of use, recency of use, and aging. The frequent activation of a node results in an increase in the rate and amount of priming that is transmitted across the connections of that node. Connections between less frequently activated nodes weaken with time, making the transmission of priming less efficient. This factor accounts for the frequency effects commonly found in speech error corpora: word and phoneme substitutions occur more often among low frequency items than among high frequency items (e.g., Stemberger, 1985, Stemberger & MacWhinney, 1986). Furthermore, object naming is faster for high frequency items than low frequency items (e.g, Oldfield & Wingfield, 1965), and TOT states tend to occur more often for low frequency than high frequency words (e.g, Burke et al., 1991). Over time, connection strength between nodes decreases. If the connections become too weak, transmission of priming becomes less efficient. The longer a node has been “inactive,” the more decay occurs to the connections of that node. Evidence for this factor comes from a diary study by Burke et al. (1991) in which young and older participants recorded naturally occurring TOT states. Burke et al. (1991) found that a TOT state would more likely occur for a proper name the longer the duration since that acquaintance was contacted. Finally, age weakens the strength of connections within the entire network (MacKay & Burke, 1990), reducing the rate and amount of priming being transmitted. This factor accounts for the general slowing of cognitive processes often associated with aging (e.g., Salthouse, 1985) and also the increased rate of TOT states seen for older compared with younger adults (Burke et al., 1991). The neighborhood density effects observed by Harley and Bown (1998)—more TOTs for words with sparse neighborhoods—can also be accounted for within NST if one views neighborhood density as a frequency effect among the phonological constituents of a word. For example, a word with a dense neighborhood has many similar sounding neighbors, and, therefore, is comprised of segments that are common, or very frequent in the language. A word with a sparse neighborhood has fewer similar sounding neighbors, and, therefore is comprised of segments that are less frequent in the language. (See Vitevitch, Luce, Pisoni, & Auer (1999) for evidence of a positive correlation between neighborhood density and the frequency of segments and sequences of segments, also know as phonotactic probability.) Phonemes (represented by phonological nodes) that constitute words with dense neighborhoods receive more priming than phonological nodes that constitute words with sparse neighborhoods. The greater amount and rate of priming received by phonological nodes that constitute words in dense neighborhoods further strengthens those connections, whereas the connections between phonological nodes and words with sparse neighborhoods become weaker over time. Thus, the phonological information associated with a word with a dense neighborhood is more efficiently retrieved than the phonological information associated with a word with a sparse neighborhood. When the number of similar sounding words (i.e., neighborhood density) is viewed in terms of sub-lexical constituents, NST can also account for the neighborhood density effects observed by Harley and Bown (1998) without invoking additional feedback mechanisms between lexemes and lemmas. NST, Neighborhoods, and Aging In the current set of experiments, we examined the influence of neighborhood density on the number of TOT states elicited experimentally with stimuli that were controlled for word length. To VITEVITCH AND SOMMERS 138 unambiguously demonstrate that neighborhood density, independent of word frequency and word length, affects TOT states, we used monosyllabic words with a consonant-vowel-consonant syllable structure that varied in word frequency, neighborhood density, and neighborhood frequency—a variable not manipulated by Harley and Bown (1998). (Neighborhood frequency is the mean frequency of the neighbors of a given target word). Also, given that the TOT phenomenon is the inability to retrieve phonological word forms from the lexicon, we used a metric of similarity based on phonological representations rather than orthographic representations as in Harley and Bown (1998). A second goal of our investigations was to examine age differences in the effects of neighborhood density on lexicalization by eliciting TOT states from older adults in Experiment 2. Previous research (Sommers, 1996) reported that neighborhood density has greater effects on older than on younger adults. Therefore, Experiment 2 examined whether this age difference in the effects of neighborhood density on perception would have parallels in production. Work by Burke et al. (1991; see also Burke & Laver, 1990; MacKay & Burke, 1990; and Rastle & Burke, 1996) suggests that older adults have more problems than younger adults retrieving items from memory during language production. For example, older adults report more TOT states than younger adults. In Experiment 2, we examined how age modulates the effects of neighborhood density on the frequency of TOT states. In summary, we examined the influence of several lexical characteristics on speech production with a stringently controlled set of stimuli, and we also investigated how the influence of these characteristics may change over the life span by eliciting TOT states from young and older adults.