Does the Medial Temporal Lobe Bind Phonological Memories?

Raymond Knott and William Marslen-Wilson Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021

Abstract

& The medial temporal lobes play a central role in the terized by a distinctive pattern of phonological errors, where consolidation of new memories. Medial temporal lesions he recombined phonemes from the original list to form new impair episodic learning in amnesia, and disrupt vocabulary response words. These were similar to errors observed earlier acquisition. To investigate the role of consolidation processes for patients with specifically semantic deficits. Amnesic Korsak- in phonological memory and to understand where and how, in off's patients showed a similar, though much less marked, amnesia, these processes begin to fail, we reexamined pattern. We interpret the data in terms of a model of lexical phonological memory in the amnesic patient HM. While representation where temporal lobe damage disrupts the HM's word span performance was normal, his supraspan recall processes that normally bind semantic and phonological was shown to be markedly impaired, with his recall charac- representations. &

INTRODUCTION mediate serial recall performance of three patients with How do transient phonological patterns in short-term semantic dementia, a neurodegenerative condition, memory become consolidated into new long-term mem- which in the initial stages, gives rise to a profound but ories? More than 40 years of research have provided relatively circumscribed semantic impairment, particu- many elegant models of auditory±verbal short-term larly affecting knowledge of word meanings. Asked to memory (AVSTM) that explain a wealth of experimental repeat short sequences of three to four content words, findings, but they have so far offered only rather meager all the patients showed a marked superiority for repeat- insights into how new phonological representations ing words that they still appeared to ``know'' compared become established. Much recent research does never- to words that are now ``unknown''Ðas established by theless suggest that AVSTM plays a central role in their performance on tests of semantic knowledge and acquiring new phonemic forms, and new vocabulary, naming. Moreover, all patients showed a distinctive in general. Gathercole and colleagues have shown that pattern of errors in their word list recall, making numer- children's performance on AVSTM tasks correlates ous phonological approximations to the original targets highly with their vocabulary knowledge (Gathercole & (e.g., ``candle sword cotton'' ! /tñndMl/ sword /t4tMn/; 1 Baddeley, 1989; Gathercole, Willis, Emslie, & Baddeley, ``number mint rug'' ! number /rInt/ /m^g/). Phono- 1992; Gathercole & Adams, 1994; Gathercole, Service, logically related errors were also noted in a variant of the Hitch, Adams, & Martin, 1999). In experimental inves- Brown±Peterson task where subjects were asked to tigations, AVSTM performance has been shown to be a repeat single spoken words after a short (3±8 sec) filled good predictor of the rate at which new phonological delay. Subsequent studies by Knott, Patterson, and forms are acquired (Gathercole & Baddeley, 1990), and Hodges (1997, 2000) suggest that this marked pattern studies using laboratory-based learning paradigms have of phonological errors is associated with semantic and/ demonstrated that the acquisition of new phonological or lexical deficits. Evidence from the normal literature vocabulary is strongly affected by the same range of supporting this view comes from a study by Treiman and factors that affect the operation of AVSTMÐi.e., phono- Danis (1988), which reported a marked and highly logical similarity, word length, and articulatory suppres- similar pattern of phonologically related errors in unim- sion (Ellis & Sinclair, 1996; Papagno & Vallar, 1992). paired subjects who were asked to repeat sequences of If AVSTM is known to affect vocabulary acquisition, monosyllabic nonwords. By definition, nonwords lack there is also strong evidence for a reciprocal influence of lexical representations. The phonological errors made vocabulary knowledge on short-term memory. Patter- by normal subjects in nonword serial recall are, there- son, Graham, and Hodges (1994) investigated the im- fore, also explicable in terms of the absence of lexical or semantic support. Further confirmation comes from a study by Knott, Patterson, and Hodges (1999), which MRC Cognition and Brain Sciences Unit, Cambridge, UK directly compared the immediate serial recall errors of

D 2001 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 13:5, pp. 593±609

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 unimpaired speakers repeating mixed sequences of the decayed trace would only provide a partial and words and nonwords. Nonwords were found to be incomplete record of the original list, attempts at recall particularly susceptible to phonological errors and re- would be characterized by phonological approximations lated studies (Poirier and St. Aubin, 1995, 1996; Bourassa to the original targetsÐi.e., phonological errors. & Besner, 1994) point to a much closer connection Toinvestigateinmoredepththewaylong-term between AVSTM and long-term memory, particularly memory interacts with short-term memory processing, long-term linguistic representations, than is often ac- it may be instructive to look at the word list recall knowledged in the short-term memory literature. They performance of other patient populations classically also suggest that in attempting to understand the rela- regarded as having impairments to long-term memory. tionship between short-and long-term memory process- Patients with anterograde amnesia are generally re- Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 ing, an important and underexploited source of garded as having largely intact AVSTM, as measured by evidence may be the errors that subjects make in digit span (Parkinson, 1982; Warrington, 1982), but they short-term memory tasks. exhibit difficulty in establishing longer-term memories If phonological errors in short-term memory are in- (Baddeley & Warrington, 1970; Drachman & Arbit, dicative of the absence of long-term memory support, 1966). In addition to difficulties on tests of episodic or what can these errors tell us about the way long-term recognition memory, studies have also demonstrated memory effects are mediated? Knott et al. (1997) offered that amnesic patients may have profound difficulties an explanation of these phonological errors in terms of acquiring new vocabulary (Kitchener, Hodges, & Mc- an interactive activation model of speech production and Carthy, 1998; Verfaellie, Croce, & Milberg, 1995). Since comprehension (Martin & Saffran, 1997; Dell & O'Seagh- auditory±verbal span remains intact in these cases, there da, 1992). On this view, lexical and semantic representa- appears to be a failure of some subsequent process that tions directly support phonological short-term memory achieves the transition to longer-term storage. The through reciprocal interactions that are continuously densely amnesic patient, HM, offers a well-documented exchanged between lexical, semantic, and phonological example (Gabrieli, Cohen, & Corkin, 1988). HM under- levels within the language system.2 Interpreted in this went bilateral medial temporal lobectomy in 1953 to manner, phonological short-term memory essentially alleviate intractable epilepsy (Scoville & Milner, 1957). reflects the pooled storage capacities of representations While this reduced the frequency and severity of his normally used in speech production and comprehen- seizures, it left him with a profound anterograde am- sion. In lexically impaired subjects, phonological errors nesia that has been extensively investigated over the arise because of the disruption of these mutually sup- intervening 45 years. In studies of both experimental portive interlevel interactions, which normally stabilize and naturalistic learning, HM has shown minimal evi- and maintain the activity of the phonological trace.3 dence of acquiring new vocabulary, and has only ac- An alternative account of phonological errors in short- quired very meager amounts of postmorbid semantic term memory can be derived from the work of Hulme, knowledge. Gabrieli et al. (1988), however, concluded Maughan, and Brown (1991) and Hulme et al. (1997). that this was not due to any additional semantic impair- This account regards long-term memory influence on ments, as HM's premorbid semantic knowledge was short-term memory as more indirect. Whereas the mod- found to be normal. HM's oral definitions of words from el described in Knott et al. (1997) assumes that short- the WAIS-R, lexical decision responses to the same term storage involves activation of phonological, lexical, words, and recognition performance for famous names and semantic representations, Hulme et al. (1991) as- from the 1930s and 1940s (Marslen-Wilson & Teuber, sume that there is a dedicated phonological storage 1975), all fell within the normal range. A detailed retro- device, or short-term store, which represents the to- spective study by Mackay, Burke, and Stewart (1998a, be-recalled list without recourse to lexical knowledge. 1998b) has, however, offered evidence that HM's seman- The phonological contents of this store are subject to tic processing may, in certain respects, not be as intact as rapid decay unless periodically rehearsed, and lexical earlier studies have suggested. In particular, HM was influences come into play when the trace is retrieved significantly poorer than controls at detecting sentential from the store. At this point, knowledge about the ambiguity or describing it. Mackay et al., (1998a) also phonological structure of words within the lexicon is report that HM's speech production was rated as less used to reconstruct the trace to its original sequence of clear, less comprehensible, and less coherent than age- items, a process likened to the pattern-completion and intelligence-matched controls. Mackay et al., attrib- property of distributed networks. Hulme et al. (1991, utes these difficulties to HM's medial temporal lobe 1997) do not give detailed consideration to error types lesion, consistent with structural MRI evidence indicat- in their account of reconstruction. Nevertheless, the ing that HM's surgery left cortical regions outside the details they provide allow us to infer how phonological vicinity of the medial temporal lobe intact (Corkin et al., errors might arise. If the lexical reconstruction process 1997). Notably, however, the controls used in Mackay were to fail, a subject's recall would be left largely et al.'s study were not on the same medication as HM. It dependent on the degraded phonological trace. Since may, therefore, be hard to determine how far HM's

594 Journal of Cognitive Neuroscience Volume 13, Number 5

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 relatively mild language problems represent nonspecific poral lobe amnesia. Prolonged alcoholism culminating effects of his medication. We will also report a study that clinically in Korsakoff's syndrome frequently gives rise to shows that despite the difficulties noted by Mackay et al., a second type of amnesia with a distinct neuropatho- HM's access to lexical representations continued to logical focus around midline diencephalic structures function efficiently. including the mammillary bodies and the medial thala- The focus of the article will be on word serial recall mus. Many Korsakoff patients have been found to show data collected for HM during the early 1970s, nearly 20 marked impairments on the Brown±Peterson short-term years after the surgical intervention that led to his forgetting task (Leng & Parkin, 1989; Cermak & Butters, amnesia. In the context of the preceding discussion, 1973). Drachman and Arbit (1966) also found poor what makes these previously unreported data of partic- performance among amnesic Korsakoff patients on a Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 ular interest is that, like more recently described cases of supraspan learning task. On the basis of these studies we semantic dementia (Knott et al., 1997; Patterson et al., might expect that, like HM, Korsakoff amnesics would 1994), HM exhibits a marked pattern of phonological perform poorly on a word serial recall task probing just errors in his word serial recall. However, unlike the beyond span length. What is of particular interest, how- phonological errors observed in the serial recall of ever, is whether Korsakoff's amnesics will show a pattern semantic dementia cases, HM's phonological errors can- of phonological errors, which, for HM, we will argue, is not be attributed to direct impairments to neocortical indicative of loss of long-term memory support. language representations. Instead, we will argue that they reflect the failure of the cognitive mechanism that Previous Studies of HM's Lesion, Language normally dynamically links together representations in Processing, and Verbal Short-Term Memory short-and long-term memory. HM's errors, therefore, provide us with an important insight into how long-term HM's medial temporal lobe resection has been described memory influences on short-term memory are mediated. by Scoville and Milner (1957). Corkin et al. (1997) provide a more precise characterization of the temporal lobe lesion based on the results of modern magnetic Phonological Errors as an Indicator of Loss of resonance imaging. We will follow Corkin et al.'s de- Long-Term Memory Support scription of the lesion. Investigations with semantic dementia patients have In studies of anterograde amnesia, HM's deficit has suggested that loss of support from long-term memory been most widely identified with bilateral excision of the representations, particularly semantic and lexical repre- hippocampus. As the MRI study of Corkin et al. (1997) sentations, leads to a marked pattern of phonological underlines, however, HM's medial temporal lobe lesion errors on short-term memory tasks (Knott et al., 1997; goes beyond damage to the hippocampal formation. Patterson et al., 1994). Whereas normal speakers mis- Approximately half of the intraventricular portion of order or omit words in immediate serial recall tasks, the hippocampal formation was destroyed bilaterally, semantic dementia cases may produce few if any of these but also the entire entorhinal cortex, the amygdaloid errors, instead exhibiting an overwhelming pattern of complex, the medial polar cortex, and portions of the phonological errors.4 Given that both HM and semantic parahippocampal gyrus (Corkin et al., 1997). The rostral dementia cases have both sustained what may be broadly portions of the perirhinal cortex were removed but described as long-term memory deficits, the occurrence some of the caudal extent remained intact. The fornix of this similar pattern of errors offers potentially impor- was spared. MRI scanning also showed that HM's surgery tant clues to the relationship between short-term phono- had not led to any additional major cortical damage. logical representation and long-term memory. Whereas There was no damage to lateral temporal cortex other in semantic dementia long-term memory impairment than at the temporal pole. HM's frontal, parietal, and appears to arise from direct damage to semantic repre- occipital cortices were all of a generally normal appear- sentations, in anterograde amnesia the primary deficit ance consistent with HM's age (66 years). There was no appears to be to a process that consolidates new episo- evidence of orbital or ventromedial frontal damage dic, semantic, or phonological memories (McClelland, arising from surgical elevation of these areas. Parts of McNaughton, & O'Reilly, 1995; Alvarez & Squire, 1994). A dorsolateral prefrontal cortex exhibited ``somewhat parallel pattern of phonological errors in semantic de- prominent'' sulcal spaces but it was not clear whether mentia and amnesia, which in both cases is attributable to this could be attributed to HM's surgery. There was, loss of long term memory support, would suggest an however, marked and diffuse atrophy of the cerebellum impairment to a single process which, in both syndromes, bilaterally, most likely arising from prolonged usage of helps to consolidate new long-term memories. antiepileptic medication. Classically, amnesia has been associated with two HM's neuropsychological and amnesic profile has distinct neuropathological syndromes. HM exemplifies been described extensively in a range of articles (see, the first type arising from bilateral damage to the hippo- e.g., Corkin, 1984b; Milner, Corkin, & Teuber, 1968, for campal complex and surrounding corticesÐmedial tem- reviews). When assessed in 1962 on Form II of the

Knott and Marslen-Wilson 595

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 Wechsler Intelligence Scale, he achieved a verbal IQ of as Milner et al. (1968) note it is higher than before his 109 and a performance IQ of 125. By comparison, his operation when he was having frequent seizures.5 HM's score of 64 on the Wechsler Memory Scale was grossly AVSTM has also been assessed by other means. Wick- impaired. Near the time of the span assessments de- elgren (1968) examined HM's short-term memory using a scribed below (1972), HM's scores on the Wechsler ``probe-digit'' task. Lists of itemsÐsingle-digit numbers, Intelligence Scale were almost unchanged (verbal IQ, three-digit numbers, or pure tonesÐwere presented 107; performance IQ, 126). auditorily. A probe item was presented immediately after the list and HM's task was to say whether the probe item had appeared in the previous list or not. For the single-

HM's Language Processing Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 digit version, lists of eight digits were presented at a rate HM's language skills have been widely reported in ear- of three items per second. For the three-digit version, lier studies to be unimpaired. Lackner (1974) studied lists were either five or seven items presented at a rate of HM's ability to identify lexical and structural ambiguities one three-digit number per second. For the tone version, in 10 word auditory sentences, containing major phrasal the first tone and the probe tone were separated by a or clausal boundaries. Asked to repeat the 10 sentences single interference tone of variable duration. On the basis immediately, HM correctly repeated 49/50, though this of these experiments, Wickelgren concluded that the score dropped to 43/50 when he was also asked to rate of decay of HM's short-term memory trace was well locate a click sound occurring within the sentence. within the normal range. Lackner concluded that HM's speech processing was The contrast between HM's normal digit span and his essentially normal. grossly impaired learning was demonstrated dramatically In a visual lexical decision task devised by Gabrieli et by Drachman and Arbit (1966). HM was asked to repeat al. (1988), HM scored equivalently to controls for pre- immediately strings of digits of different lengths using a 1950s words, but was more than two standard deviations presentation rate of one digit per second. If a list was below the normal mean for post-1950s words which had repeated accurately, a string one digit longer was pre- entered the language subsequent to HM's operation. sented. If it was repeated incorrectly the same string was When HM was asked to provide definitions of the same presented again until it could be repeated correctly. words to assess his semantic knowledge, his scores were Although HM showed a normal span of six digits on again comparable to controls for pre-1950s words, but this task, he was entirely unable to recall a sequence one he showed a marked impairment for post-1950s words. digit longer than his span, even after 25 repetitions of In a recent study, employing a word-stem completion the same seven-digit list. This clearly suggests that task (Postle & Corkin, 1998), HM was asked to complete memory processes that support HM's recall immediately three-letter word fragments with the first word that beyond span are inoperative. came to mind. HM showed normal priming for pre- 1953 words. For words that entered the language after Objectives of the Present Study 1965, 12 years after his operation, his priming score did not differ from zero, underlining HM's highly impover- Our study has two main aims; firstly to reexamine HM's ished postmorbid vocabulary acquisition. short-term memory performance on the word serial Two retrospective studies by Mackay et al. (1998a, recall task. Like earlier authors we will conclude that, 1998b) have recently reevaluated earlier findings (Lack- when assessed on tasks such as digit span, HM's AVSTM ner, 1974; Corkin, 1984a; Mackay, 1972; Marslen-Wilson, is unimpaired. In more contemporary terms, we will say 1970) and suggest that HM's language processing abil- that HM's ``phonological'' short-term memory, under- ities, even for premorbidly acquired vocabulary, may not stood as the transient phonological code emphasized by be as fully intact as previous studies have assumed. In Baddeley and many other authors, remains intact (Mar- particular, Mackay et al. (1998a, 1998b) identified some tin & Lesch, 1996; Burgess & Hitch, 1992; Gathercole & mild problems with production and comprehension of Baddeley, 1993; Hulme et al., 1991; Baddeley & Hitch, novel sentences. Taken together, however, the full range 1974). The second aim of the study has been to inves- of studies suggest that HM's language representations tigate AVSTM in anterograde amnesia at list lengths just for premorbidly acquired vocabulary, and the language beyond span (in HM and a group of span-matched processing mechanisms that operate on them, are not Korsakoff amnesics). Assessing short-term memory per- markedly impaired. formance just beyond span emphasizes the functioning of memory processes, which on contemporary views of short-term memory, provide additional support for the HM's AVSTM Performance phonological code (Poirier & St. Aubin, 1995; Hulme et Since his surgical intervention, HM has been consistently al., 1991, 1997). These additional memory processes assessed to have a forward span of six digits (Marslen- have been shown to be impaired in patient populations Wilson & Teuber, 1975; Drachman & Arbit, 1966). such as semantic dementia, which have sustained pri- Although this may be somewhat low for his intelligence, mary impairments to long-term semantic representa-

596 Journal of Cognitive Neuroscience Volume 13, Number 5

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 tions (Knott et al., 1997). We will demonstrate that when 1 tested just beyond span length, both HM and amnesic HM Korsakoff patients are markedly impaired in their recall. 0.8 Normal More significantly, their recall is characterized by a 0.6 Controls pattern of phonologically related errors, highly similar Head-Injured to those seen in semantically impaired subjects. The 0.4 Controls Korsakoff's patients implications of these findings will be discussed in terms 0.2 of a model in which short-term phonological represen-

tation are supplemented by access to additional long- Proportion of total words correct 0 term lexical and semantic representations. We will con- 12345 Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 Serial position of target clude that owing to his medial temporal lobe impair- ment, HM has lost the ability to dynamically link these two types of representation together.

RESULTS 1 HM In addition to HM, three groups of participants took part 0.8

in the studyÐnormal controls, head-injured controls, 0.6 Normal Controls and Korsakoff amnesic patients. All participants were Head-Injured initially administered a word serial recall task to establish 0.4 Controls Korsakoff's their span level. Participants were asked to repeat blocks 0.2 patients of five lists each consisting of sequences of auditorily 0 presented monosyllabic words. Span level was defined Proportion of total words correct 123456 as the list length at which participants recalled at least Serial position of target 80% of the lists in each block correctly. Based on this criterion, HM's span level was assessed as four mono- syllabic content words. From each group (normal con- Figure 1. Proportion of words recalled as a function of serial position trols, head-injured controls, and Korsakoff amnesics) for List 1 (top) and List 2 (bottom). participants were then selected who matched HM's span level. To assess short-term memory performance just 3.29, MSE = 16.4, p < .05; List 2: F(2,40) = 3.57, MSE = beyond span level, each group was then given two 22.1, p < .05, and serial position, List 1: F(4,160) = 36.5, further immediate serial recall tasks using sequences of MSE = 3.94, p < .0001; List 2: F(5,200) = 47.9, MSE = monosyllabic words that exceeded measured span by 7.11, p < .0001, but no interaction between group and either one or two words. We will label these tasks as serial position, List 1: F(8,160) = 0.35, MSE = 3.94, p > List 1, five-word sequences (span + 1), and List 2, six- .05; List 2: F(10,200) = 0.92, MSE = 7.11, p > .05. A priori word sequences (span + 2). tests showed that the Korsakoff amnesics recalled sig- Table 1 shows the total number of words correctly nificantly fewer words than controls, List 1: t(41) = 2.07, recalled at each supraspan list length for HM, and the p < .05; List 2: t(41) = 2.65, p < .01. HM recalled fewer corresponding means for each of the three groups still (Table 1).6 The two control groups did not differ, List (normal controls, head-injured controls, and Korsakoff 1: t(33) = 0.99, p > .05; List 2: t(33) = À0.11, p > .05. amnesics). Figure 1 illustrates word recall scores as a Each of the two impaired groups (head-injured con- function of serial position. A two-way (List  Group) trols and Korsakoff amnesics) shows a normal serial ANOVA indicated that participants found List 1, the recall curve displaying the usual prominent primacy shorter list, to be significantly easier to recall than List and recency portions (Figure 1). Although in terms of 2, F(1,40) = 268, MSE = 0.51, p < .001. For each list total words recalled, HM scores more poorly than either length, two-way (Group  Serial Position) mixed AN- of the control groups, he shows a relatively normal serial OVAs revealed main effects of group, List 1: F(2,40) = position curve with both primacy and recency portions visible. This is consistent with the conclusion that HM's Table 1. Mean Words Correctly Recalled (Standard Deviations AVSTMÐunderstood as the short-term storage capacity in Parentheses) that underlies span performanceÐremains normal. For comparison, Martin and Lesch (1996) and Saffran and Normal Head-Injured Korsakoff Martin (1990) report cases with impaired span perform- Controls Controls Amnesics ance, which show markedly abnormal serial position (n = 8) (n = 27) (n = 8) HM curves on similar immediate serial recall tasks. List 1 79.6 (4.98) 76.2 (9.53) 69.6 (10.3) 62.0 Aphasic patients may also exhibit strong frequency and/or semantic effects in their immediate serial recall List 2 72.13 (11.8) 72.7 (11.8) 60.7 (10.2) 46.0 (Knott et al., 1997, 2000). HM's recall was uninfluenced

Knott and Marslen-Wilson 597

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 Table 2. Mean Number of Phonemes Recalled in Each List further measures of retention were obtainedÐthe num- (Standard Deviations in Parentheses) ber of phonemes that were recalled outside their orig- Normal Head-Injured Korsakoff inal word targets, and the overall number of target Controls Controls Amnesics phonemes recalled (either in their correct word targets, (n = 8) (n = 27) (n = 8) HM or in phonologically related errors). List 1 289 (11.1) 284 (26.7) 276 (19.5) 297.0 Table 2 shows the total number of phonemes re- tained. Each figure was computed by adding together List 2 294 (41.2) 239 (34.8) 292.0 (27.0) 302.0 the number of phonemes recalled in target words (e.g., recalling ``dog'' as /d4:g/) and the number of phonemes by frequency at either list length, List 1: mean frequency that were recombined into new words (e.g., recalling Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 of corrects = 50.4, mean frequency of incorrects = 75.7, ``dog'' as /l4:g/Ðtwo phonemes recalled). Examples of t(47) = 1.49, p > .05; List 2: mean frequency of corrects HM's list recall responses are shown in Table 3. HM = 91.1, mean frequency of incorrects = 97.0, t(93) = frequently produced responses that were phonologically 0.19, p > .05. There was however a trend toward better related to a list target, retaining one or more the target retention of higher imageability words for both sets of phonemes, e.g., ``chance'' ! /txñnt/, ``hook'' ! /haIk/. lists, List 1: mean imageability rating of corrects = 541, Other errors suggest recombinations of phonemes from mean imageability of incorrects = 504, t(44) = 1.86, p < different list targets, ``fetch plant' ! /flñntx/. As we will .10; List 2: mean imageability rating of corrects = 536, demonstrate below, it is the frequency of HM's phono- mean imageability of incorrects = 507, t(100) = 1.93, p logical errors that distinguishes him most clearly from > .10.7 Over the restricted range of word lengths used the controls. While the controls also made phonological (one-syllable items, ranging from two to six phonemes in related errors in their recall, they made them at a much length), length in phonemes did not influence recall, List lower rate. When a control was unable to correctly recall 1: mean phoneme length of corrects = 3.41, mean an item, a null would commonly be marked where the phoneme length of incorrects = 3.43, t(70) = 0.13, p control failed to produce any explicit response. It is > .05; List 2: mean phoneme length of corrects = 3.44, noteworthy that while HM produced many more pho- mean phoneme length of incorrects = 3.37, t(111) = nemic errors, he also made fewer null responses, sug- 0.40, p > .05. gesting a difficulty in monitoring the correctness of his Since the list recall task was scored on a free-recall responses, List 1: mean number of null responses for criterion, HM's much poorer recall compared to con- normal controls = 6.63, range = 3±10, HM = 1; List 2: trols must be accounted for in terms of item errorsÐ mean number of null responses for normal controls = either omissions or intrusions. When a word was not 25.9, range = 13±40, HM = 7). recalled correctly, control subjects would occasionally In the vast majority of HM's phonological errors, a report an intrusion, either a word unrelated to any list retained phoneme was recalled at a syllabic position target, or a phonologically similar word that retained similar to its original syllabic position in the target word. phonemes or segments from the misremembered item, Thus, 55/59 consonants, which were originally in onset e.g., ``force'' ! /fi:rt/. Errors that retained phonemes position in the target, were recalled in a recombination from the list, which were not recalled correctly else- error as a constituent of an onset. Similarly, 51/53 where, were marked to record the number of retained consonants, originally constituents of codas, were re- segments. In the example above, three retained seg- called as constituent of codas. Despite this general ments would be tallied. By following this procedure, two adherence to syllable position constraints, retention of

Table 3. Examples of HM's Phonemic Errors Target List HM's Response List 1 tire lamp fetch plant heel tire /left/ /flñntx/ /pi:Ml/ tire chance bomb score fist tool /txñnt/ bomb /skIp/ /f4:rs/ /txu/ league snow joy mat pinch league /dkoP//miI/ /pñtx/±

List 2 ship web life tramp cave roll ship /waIf/ /trIp/ /kñ^k/ /rñt/ /xoPl/ edge fruit play latch dime style edge fruit latch /naIm/ /deI/ style night hook chief size marsh glue night /haIk/ /nri:/ /bPk/ /dkg:R/ /krñx/

598 Journal of Cognitive Neuroscience Volume 13, Number 5

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 Table 4. Recombined Phonemes as a Percentage of Total 1 HM Phonemes Recalled (Standard Deviations in Parentheses)

0.9 Normal Controls Normal Head-Injured Korsakoff

Head-Injured Controls Controls Amnesics 0.8 Controls (n = 8) (n = 27) (n = 8) HM Korsakoff's patients List 1 8.15 (3.03) 10.2 (3.88) 16.9 (6.7) 31.3 correctly recalled 0.7

Proportion of total phonemes List 2 16.0 (6.05) 17.1 (6.4) 28.1 (7.1) 47.7 0.6 12345 Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 Serial position of target errors, a final measure was calculated. This quantified the percentage of phonemes that were not recalled in their original words, but which were produced either through recombination into other words (real words or 1 nonsense words), or as fragments of the original words. 0.9 HM This measure is summarized in Table 4 and in Figure 3. 0.8 Here, the clearest differences between subjects Normal Controls emerged. Two-way (Group  Serial position) ANOVAs 0.7 Head-Injured indicated main effects of group, List 1: F(2,40) = 9.56, 0.6 Controls

correctly recalled Korsakoff's MSE = .011, p < .001; List 2: F(2, 40) = 10.7, MSE = 0.31, 0.5 patients p < .001. A priori tests showed that the Korsakoff Proportion of total phonemes 0.4 amnesics produced a significantly larger proportion of 123456 phonological errors than either control group, List 1: Serial position of target Korsakoff patients > normal controls, t(14) = À3.38, p < .01, Korsakoff patient > head-injured controls t(33) = Figure 2. Proportion of phonemes recalled as a function of serial À3.22, p < .01; List 2: Korsakoff patients > normal position for List 1 (top) and List 2 (bottom). controls, t(14) = À3.54, p < .01, Korsakoff patients > head-injured controls, t(33) = À3.94, p < .001. Mean- while, HM's score was more than three standard devia- syllabic position, or role within a syllable constituent was not demanded as a condition for marking a phoneme as retained. Thus, in the response ``fetch plant'' ! /flñntx/, HM where five phonemes would marked as retained, the /f/ 0.6 Normal sound was marked even though it was now part of an Controls 8 onset cluster, rather than a singleton. Occasionally 0.4 Head-Injured HM's phonologically related errors resulted in what Controls Korsakoff's were apparently nonsense words, ``snow joy'' ! /dkoP/ patients recalled 0.2 /miI/, but it is notable that the great majority of his intrusion errors preserved lexical status. Excluding in-

trusion errors that were simply repetitions of earlier list Phoneme fragments / Phonemes 0 targets, fully 84% of HM's phonologically related errors 12345 preserved lexical status. Serial position of target Figure 2 shows HM's responses expressed in terms of the proportion of phonemes retained as a function of serial position. Group  Serial Position ANOVAs re- vealed no differences between the three groups on this 1 HM measure, List 1: F(2,40) = 0.21, MSE = .05, p > .05; List 0.8 2: F(2,40) = 0.83, MSE = .025, p > .05. Table 2 shows Normal 0.6 Controls that HM's phoneme recall scores also lie close to the Head-Injured mean for normal controls (both within one standard 0.4 Controls deviation and slightly above). This is in marked contrast recalled Korsakoff's 0.2 patients to his word recall scores in Table 1. 0

Despite the overall similarity in the level of phoneme Phoneme fragments / Phonemes recall for HM and controls, Figure 2 suggests that the 123456 recency effect visible in the control data may be attenu- Serial position of target ated in HM for both Lists 1 and 2. In order to compare more formally with controls the Figure 3. Phoneme fragments as a proportion of phonemes recalled rate at which HM was producing phonologically related for List 1 (top) and List 2 (bottom).

Knott and Marslen-Wilson 599

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 Table 5. Proportion of Words Recalled for HM and Re- control mean by more than four standard deviations for matched Controls for Lists 1 and 2 List 1 and more than three standard deviations for List 2. Fragments as a Thus, despite rematching on the supraspan tasks to Proportion of Proportion of equate the total number of words recalled, HM contin- Words Recalled Phonemes Recalled ues to produce substantially more phonological errors List 1 List 2 List 1 List 2 than the controls. Do the Korsakoff amnesics also continue to show a HM 0.62 0.38 0.32 0.51 higher rate of phonological error once overall number of words recalled is taken into account? To answer this Rematched 0.62 0.46 0.15 0.24 question, and to investigate the relationship between Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 controls (‹0.26) (‹0.09) (‹0.07) (‹0.16) phonological error production and recall level in the Means shown for rematched controls with 95% confidence intervals control and Korsakoff subjects, we constructed scatter- (‹ 2 SDs, in parentheses). plots. Figure 4 shows the percentage of phoneme frag- ments plotted against total errors for the control tions beyond the Korsakoff patients' score at each list subjects, the Korsakoff amnesics, and HM. For List 1, length (Table 4). Thus, the difference between the word there is a significant correlation between the total and phoneme scores in Tables 1 and 2 reflects the greater number of errors that the controls make (i.e., failures incidence in the Korsakoff group, and especially in HM, to recall the target) and the number of phoneme frag- of phonemic recombination errors. Almost half of HM's ments they produce (multiple correlation coefficient, total phoneme score on the longer list (List 2) derives R = .45, t(34) = 2.90, p < .01). For List 2, the controls from phonemes recalled outside their original list targets. show a trend in the same direction (multiple correlation coefficient, R = .29, t(34) = 1.77, p < .10). For both lists, this indicates a tendency for control subjects to produce HM's Phonological Error Rate Compared to more phonemically related errors as the total number of Controls Rematched to HM for Supraspan errors they make increasesÐthat is, as they are pushed Performance further beyond their normal span length. Regression In the analysis described up to this point, HM's word lines relating the number of phoneme fragments to span has been taken as four, based on the criterion that the total number of errors are shown for the control span was the maximum list length at which a subject subjects in Figure 4 with 99% confidence limits (for List

produced at least 80% (i.e., 4/5 lists) within a block 1, the regression coefficients are b0 = 15.2, b1 = 0.57; for correctly. It might be claimed, however, that this is not a List 2, b0 = 0.43, b1 = 29.9). The tendency for normal sufficiently sensitive measure of span, and that HM's poorer performance on the supraspan tasks arose be- cause his word span was in reality slightly below con- 100 Head-injured and normal trols. On this basis, it could be argued that the large 80 99% confidence intervals controls number of phonologically related errors in HM's supra- 60 Korsakoff span recall was not, in fact, abnormal, but arose simply amnesics HM because he was being asked to recall lists which, in 40 comparison to controls, were even further beyond his 20 Total phoneme fragments normal span. In order to address this potential claim, we 0 carried out a further analysis. Using HM's supraspan 01020304050

word recall scores from Table 1, he was rematched to Total errors a subgroup of five head-injured subjects from the orig- inal span-four control groups. These controls were selected as those whose overall word recall scores, when 160 Head-injured and normal summed across the two list lengths, were closest to 99% confidence intervals controls 120 HM's. They were also the lowest scoring subjects within Korsakoff amnesics the head-injured group; no normal controls were in- 80 HM cluded in the rematched group because all exceeded HM's word recall on the supraspan lists. 40

Mean word recall scores for the rematched controls Total phoneme fragments are shown in Table 5. For List 1, HM's word recall score 0 0 20406080100 is the same as the control mean. For List 2, HM's score Total errors lies within two standard deviations of the control mean. Table 5 also includes the mean proportion of phonemes that were recalled by controls outside of their original Figure 4. Scatterplot of phoneme fragments against total errors for targets. For HM, this percentage exceeds the overall List 1 (top) and List 2 (bottom).

600 Journal of Cognitive Neuroscience Volume 13, Number 5

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 subjects to make more phonological errors as span here were a subset of a larger set of materials developed to length is exceeded is, however, clearly insufficient to probe the on-line relationship between aspects of lexical account for HM's phonological error rate. As the figure processing and sentence-level syntactic and semantic illustrates for both list lengths, HM's phonological error constraints (Marslen-Wilson & Tyler, 1975, 1980). HM score substantially exceeds the 99% confidence limits. By was asked to monitor for word-targets in two types of comparison, all but one of the Korsakoff amnesics falls sentential context, normal prose, and anomalous prose. within the 99% confidence limits. To a large extent the In normal prose, he heard a passage such as the following: increased rate of phonologically related errors produced by the Korsakoff amnesics on the supraspan lists can be The church was broken into last night. Some thieves attributed to their inferior word recall scores at these stole most of the lead off the roof. Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 supraspan lengths. We return to this important differ- ence between HM and the Korsakoff patients in the In anomalous prose the sentences were syntactically Discussion. still interpretable but had no coherent semantic inter- pretation: DISCUSSION The power was located into great water. No buns Assessed on lists of either words or digits, HM's span puzzle some in the lead off the text. performance is comparable to normal subjects, suggest- ing that his immediate memory, or in more contempo- For each type of prose context, three monitoring rary terms his phonological short-term memory, remains conditions were used (identity, category, and rhyme). intact. On lists just beyond span length, however, differ- In identity monitoring, HM was asked to press a button ences between HM and control subjects rapidly begin to as rapidly as possible when he heard a prespecified emerge. HM's supraspan recall of complete words is wordÐfor the examples above this would be ``lead.'' poorer than all of the normal control subjects, and all In category monitoring, he had to respond when he but five of the head-injured controls. His responses are heard a word that belonged to a specified semantic characterized overwhelmingly by acoustically related er- categoryÐin this case ``a kind of metal.'' In rhyme rors, which preserve much of the phonemic content of monitoring, HM was asked to listen for a word that the target list. Indeed, despite his errors, in terms of rhymed with a given wordÐhere the cue word would be phonemic content HM recalls the same proportion of ``bread.'' The relative speed of processing in the three target phonemes as the control subjects, even while the conditions, and the effects of the presence or absence of control subjects recall many more of the original words. semantic constraints in the prose context, allow us to Simply put, HM recalls the phonemic content just as evaluate HM's on-line abilities in this domain. This was accurately as the controls, but recalls much of it outside intended as a preliminary investigation and we did not the original word targets as phonemically related errors. have available an age-and education-matched control What leads to this marked pattern of phonological group specifically designed for HM. Instead, we will use errors in HM's supraspan recall? Given HM's normal span the data for the young adults (University of Chicago level performance, it seems unlikely that it can be ex- undergraduates) tested in the original experiments plained as a simple capacity limitation in short-term (Marslen-Wilson & Tyler, 1975). HM's performance, as memory. In the semantic dementia cases reported by reported in Table 6, is nonetheless broadly comparable Knott et al. (1997, 2000) and Patterson et al. (1994), to that of the normal controls. He is just as fast as the phonological errors in word serial recall were found to average undergraduate in the normal prose identity be associated with lexical and semantic impairments. monitoring condition, indicating the efficiency with Mackay et al. (1998a, 1998b) have recently claimed that HM may have sustained subtle high-level impairments in Table 6. Mean Monitoring Reaction Times for HM and language comprehension and production that particu- Unimpaired Controls larly affect his ability to comprehend or produce multi- word utterances. For example, HM had difficulty Monitoring Task identifying and explaining the different meanings of sen- tences that were structurally ambiguous. Other evidence, Prose Context Identical Rhyme Category Mean however, suggests that HM's on-line semantic processing of sentences and individual word meaningsÐmore di- Normal prose HM 271 532 472 425 rectly relevant to the claims we are making here about Controls 273 419 428 373 immediate recall of lists of auditory wordsÐremained Anomalous prose HM 428 505 591 508 substantially intact. This evidence comes from additional testing sessions, also in the early and mid-1970s, where Controls 331 463 528 440 HM was asked to monitor for target words presented Controls were 45 students at the University of Chicago. Overall auditorily in a variety of sentential contexts. The stimuli standard error for controls = 8.75 msec.

Knott and Marslen-Wilson 601

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 which he is able to process basic information about to the nonwords, which by definition lacked full lexical lexical identity and to make use of the cues provided by representations. a normal sentential environment. His performance is In summary, we have enumerated evidence for pho- much worse in the anomalous prose condition, where nological errors resulting from lack of long-term memory the absence of semantic and pragmatic constraints slows support in three contexts: in HM, who cannot establish down his responses by over 150 msec. new long-term memories; in semantically impaired cases, HM's performance is also quite normal in the category who have lost long-term memories; and in normal sub- monitoring condition, which places a premium on rapid jects repeating nonwords, who, because of the choice of semantic analysis of targets in their sentence contexts. nonwords, cannot exploit long-term memories to sup- His responses in the normal prose condition are only 44 port their recall. HM's data provide important new Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 msec slower than the young controls, as well as being evidence by showing that the occurrence of these errors highly accurate (controls make essentially no errors in need not depend on the absence of, or direct impair- this task; HM's error rate was 6%).9 When the prose type ment to, appropriate long-term memory representations. is switched from normal to anomalous prose, reducing the available semantic cues in the sentential context A Model of Phonological Binding leading up to the target word, HM shows a sharp increase in category monitoring latencies compared to To try to account for the parallel pattern of phonological rhyme monitoring. A parallel effect is seen in the con- errors in these three contexts, we will first introduce a trols (Table 6). This selective increase in category mon- model of short-term memory by Hulme et al. (1991, itoring latency illustrates HM's sensitivity to semantic 1997), which was explicitly developed to account for information and his ability to access and interpret it long-term memory effects in the immediate serial recall rapidly and on-line. Taken together, these findings argue task. According to Hulme et al., when a partially decayed against a significant semantic impairment in HM during trace is retrieved from phonological storage, it undergoes the time period of the supraspan memory tests reported a lexically mediated process of redintegration, which here, and clearly distinguish him from semantic demen- reconstructs it using stored lexical knowledge. Where tia patients whose semantic impairment has previously reconstruction is successful, this restores the degraded been held to be the source of phonological errors in trace to its original sequence of lexical items (Hulme et word serial recall (Patterson et al., 1994) al., 1991, 1997). On this account, the degraded phono- Even in semantic dementia, moderate semantic im- logical trace can be considered as a sequence of phonetic pairments do not inevitably lead to a pronounced pat- or phonemic cues that at recall can be used to access tern of phonological errors in word serial recall. Knott et matching lexical representations, which then respecify al. (1997) reported a patient who showed only minimal the original target. As the trace becomes more degraded evidence of phonological errors in word serial recall, over time, and as the list length increases beyond span, even for words that he no longer comprehended. This this reconstruction process can be expected to become patient nevertheless exhibited much more profound more important for ensuring accurate recall. comprehension and production difficulties than HM. Hulme et al. (1991, 1997) were not able to specify the Notably, Mackay et al. (1998a, 1998b) do not attribute reconstruction process in detail, or describe the mech- HM's apparent difficulties with ambiguous sentences to anisms that implement it, though they liken it to the any ``direct'' impairment to semantic representations. pattern-completion property of distributed networks. Instead, they attribute his comprehension difficulties to One way to conceptualize reconstruction would be to his medial temporal lobe lesion. Developing their argu- regard it as a process that simply matches the contents ment in terms of Mackay's (1987) node structure theory of the phonological representations in short-term mem- (NST), they argue that producing or comprehending ory against patterns stored in the lexicon, selecting as novel utterancesÐnovel utterances being those that the recall item the best match. HM, however, clearly has HM has particular difficulty withÐrelies on the forma- some retained capacity to accomplish this. His phono- tion of new cortical±cortical links between existing rep- logically related recall errors are only rarely nonwords. resentations. According to Mackay et al. (1998a, 1998b), Even though he cannot remember the original word list establishing these links requires direct input from the correctly, his responses, therefore, indicate continued hippocampus, which in HM is unavailable. access to lexical knowledge. A simple reconstruction Although semantic impairment may not always lead to account also fails to explain why HM can recall phono- an unusual pattern of phonological errors in word serial logical information from short-term memory at the same recall, other lines of evidence nevertheless suggest that level as controls (cf. Table 2) while being unable to phonological errors are associated with either dimin- remember the words that these phonological patterns ished lexical or semantic support. Knott et al. (1999) represented only moments after hearing them. directly compared immediate serial recall errors made to Simply put, HM recalls the phonemic content of a list words and nonwords in unimpaired speakers. Partici- correctly, but is unable to remember the words that pants produced substantially more phonological errors these phonemes occurred in. What seems to be at fault

602 Journal of Cognitive Neuroscience Volume 13, Number 5

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 in HM is a process that links representations in phono- TEMPORAL & PARIETAL logical short-term memory to recently activated lexical or NEOCORTEX semantic representations. This observation suggests an Semantic and lexical representations intriguing connection between HM's errors in phono- of familiar words logical memory and his well-documented episodic mem- ory impairment. To try to capture this connection, we Fast-changing weights connecting cortex to the will turn to a view of medial temporal lobe function medial temporal lobe developed by Alvarez and Squire (1994). In Figure 5 we offer a model that will allow us to account for HM's supraspan recall errors. In terms of its general architec- Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 tural and processing assumptions, it is closely related to the model of retrograde and anterograde amnesia de- veloped by Alvarez and Squire (1994). Most saliently, it MEDIAL TEMPORAL LOBE assumes that the medial temporal lobe plays a central role in consolidating new memories. We will first briefly describe some of the background to this view. Wide-ranging evidence implicating medial temporal lobe structures in the consolidation of new memories Phonological representations comes from both neuropsychological investigations of in short-term memory anterograde and retrograde amnesia, and lesioning stud- ies with animals (see Squire, 1992, for a review). The consolidation process may take place over a timescale ranging from a few seconds to several years following the Speech input and output initial event. In the present context, we will focus on the very earliest stages of the consolidation processÐwhat Figure 5. A model of phonological binding and vocabulary acquisition. Nadel and Moscovitch (1997) term the ``cohesion'' stage. The existence of cases showing preservation of premor- semantic representations become automatically acti- bid episodic and semantic memories in the context of vated by direct connections within the language process- profound anterograde amnesia has led most investiga- ing system. Fast-changing weights that connect the tors to infer that the neocortex, rather than the medial medial temporal lobe to each of these respective neo- temporal lobe, is the ultimate repository of new long- cortical regions register this pattern of coactivation by term memories. Consolidation can, therefore, be inter- increasing in strength. This amounts to encoding a weak preted as the binding together of the many geographi- episodic memory of the event in connections between cally separate regions of neocortex that together form the medial temporal lobe and neocortex, so that when the representation of a new memory. Direct connections some fragment of the original word-form is re-presented between neocortical areas may, however, change only to the phonological level, the connections via the medial very slowly. To facilitate the process of consolidation, temporal lobe reinstantiate the original pattern of lexical Alvarez and Squire's model proposes that fast-changing activation. In this way, the network behaves as a pattern- connections to and from the medial temporal lobe associator (Alvarez & Squire, 1994). When lesioned temporarily bind together neocortical regions. Over appropriately, the model can also account for the pat- time, as direct neocortical connections strengthen, stor- tern of phonological errors in both the amnesic patients age and retrieval becomes less dependent on the medial and the semantic dementia cases. temporal lobe, and the connections to and from the When a subject attempts to retrieve a word directly medial temporal lobe eventually become redundant. from the phonological trace, particularly in a supraspan We will follow Alvarez and Squire in assuming that the task that exceeds the level of comfortable recall, the consolidation of new memories is mediated by a set of decayed phonological representation may not be suffi- fast-changing weights, to and from the medial temporal cient to support full recall. Because of HM's extensive lobe, which link together geographically distinct areas of medial temporal lobe lesion, his recall is largely reliant neocortex. These fast-changing connections provide a on this partially decayed phonological memory trace. means for linking, or binding together, coactivated re- HM's recall is, therefore, characterized by numerous gions of cortex, which then together form the basis of errors that are phonetic or phonological approximations the new memory. Two distinct levels of representation to the original target items. In unimpaired subjects, are depicted in the model in Figure 5Ðlexical/semantic however, the residual phonological information can be representations and phonological representations, both used as a cue to reactivate the corresponding lexical of which we located within the lateral portions of the representation via intact connections to and from the temporal and parietal lobes. When the subject hears a medial temporal lobe. As a consequence, phonologically word in the serial recall task, these phonological and related errors are a relatively infrequent occurrence.10

Knott and Marslen-Wilson 603

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 Semantic dementia patients have sustained damage to usually selected from among these competing candi- a different section of the reconstruction circuit. Like HM, dates. In HM this judicious selection cannot be made. they have unimpaired phonological representations, but There is one major omission from the discussion so far. their lexical and semantic representations are no longer No mention has been made of the Korsakoff patients, and functioning effectively. It seems likely that many seman- whether they fit into this scheme. The performance of the tic dementia patients will also have damage to connec- Korsakoff patients was in many respects similar to HM. tions between the medial and lateral temporal lobe that Like HM, they exhibited a normal word span, but impaired are proximal to the atrophied portions of inferior and word recall just beyond span length. As a group, and lateral temporal lobe (possibly perirhinal and parahippo- similarly to HM, they nevertheless showed an equivalent campal cortices), which are more commonly recognized level of phonemic retention to controls, producing an Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 as impaired in these patients. The consequences of this unusually large number of phonemes from the target list damage for supraspan serial recall will be similar to HM. in their errors. In many respects, therefore, their perform- In both patient groups it will impair the efficiency with ance was qualitatively similar to HM, and distinct from which the circuit can reconstruct, or complete, decayed either of the nonamnesic control groups. However, they phonological patterns. Semantic dementia patients are, still produced significantly fewer phoneme fragments in therefore, predicted to produce a similar pattern of their responses than HM. Moreover, when considered as phonological errors in their supraspan serial recall as individual cases with overall level of word recall also taken observed (Knott et al., 1997; Patterson et al., 1994). into account, only HM showed a consistently abnormal Aside from the data from these two patient groups, we rate of phoneme fragment production (Figure 4). Thus, need to consider whether there is any independent the regression analyses indicate that, for the Korsakoff justification for the claim that the medial temporal lobe patients, the frequency with which phoneme fragments binds together coactivated semantic and phonological were recalled could be accounted for in terms of their representations. Beyond the explanation of errors in overall poor word recall scores. This is consistent with our serial recall tasks, what possible reason could there be claim that the pronounced pattern of phonological errors for binding a transient representation in phonological in HM arises from damage to medial temporal lobe short-term memory to a semantic representation? We structures. Neither diencephalic lesions nor indeed retro- suggest that the solution to this puzzle is to recognize, as grade amnesia are by themselves sufficient.11 much research now indicates, that an important, possibly The connection between the phonological errors ob- primary function of phonological short-term memory is served in HM and in semantic dementia requires further to mediate vocabulary acquisition. The binding mecha- elucidation. In semantic dementia, pathology is most nism that we have proposed facilitates this process. It visible in infero-lateral regions of temporal cortex, focus- dynamically links together a phonological and a semantic ing on the temporal pole, predominantly, but not ex- representation, providing a way to rapidly associate a clusively, in the left hemisphere. In HM, white matter previously unexperienced phonological form with a tracts to and from the temporal lobe were certainly found cooccurring pattern of semantic activation. On this mod- to be damaged in Corkin et al.'s MRI study.12 Indeed, el, the early representation of the new word is highly HM's medial temporal lobe lesion goes well beyond the dependent on the fast-changing connections to and from hippocampus, subsuming portions of the adjacent ento- the medial temporal lobe. With repeated exposure to the rhinal, perirhinal, and parahippocampal cortices bilater- same item, however, slower-changing neocortical con- ally. It may be that it is additional damage to these nections between semantic and phonological represen- regions of temporal cortex that leads to HM's particularly tations also become established, such that eventually marked pattern of phonological errors. It would be coactivation of the corresponding semantic and phono- valuable to know the extent to which HM's lesion over- logical representations can be accomplished without laps, in this respect, with the pathology seen in semantic support from the medial temporal lobe. At this point, dementia. The status of the proximal entorhinal, peri- the new item of vocabulary can be considered to have a rhinal, and parahippocampal cortices in semantic de- fully established lexical representation, based on direct mentia is currently uncertain, but under investigation. neocortical connections. These neocortical connections are shown as a dotted line in Figure 5. This model may also provide some new and testable predictions about the METHODS nature of retrieval from phonological short-term memo- ry. In particular, it implies that retrieval from phonolog- Subjects ical short-term memory is a guided process. Although a In addition to HM, three further groups of male partic- decayed phonetic and phonological fragment retained in ipants took part in the experimentÐa group of amnesic the trace may be a plausible cue for a variety of different Korsakoff's patients, a group of head-injured controls, lexically represented words, the rapid weight-changes to and a group of unimpaired controls. The group data connections to the medial temporal lobe that occur at the were collected over the period 1970±1973. HM's data encoding stage ensure that, at recall, the correct word is were collected in 1972 when he was 46 years old.

604 Journal of Cognitive Neuroscience Volume 13, Number 5

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 Amnesic Korsakoff Patients and verbs in the Thorndike±Lorge word-frequency counts (Thorndike & Lorge, 1944). Two sets of lists The onset of amnesic symptoms in the Korsakoff's were constructed by selecting words pseudorandomly group (n = 12, mean age = 54.7) varied from 11 months from the pool, with the constraints that words in any to nearly 10 years in duration (mean memory quotient = one list should not be obviously associated, or phoneti- 71.7, range = 62±93; mean verbal IQ = 104.1, range = cally similar. Lists were prerecorded for the experiment 95±123; mean performance IQ = 98, range = 86±116). by a native speaker of American English. All subjects were well past the initial florid state typical of the onset of the Korsakoff's syndrome. Their medical

histories recorded that they had each been addictively Procedure Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 alcoholic for a prolonged period ranging from 15 to 34 Two successive sets of 20 lists, each composed of years, prior to the onset of their amnesia. At some point, monosyllabic words, were presented to the subjects in they had all presented with signs of acute Wernicke± each group. List 1 consisted of strings one word longer Korsakoff encephalopathy, i.e., either oculomotor dis- than the subject's span. List 2 consisted of lists two turbances or ataxia, or both combined with peripheral words beyond span. Lists were spoken to the subjects at neuropathies. At the time of testing, however, their the rate of one word per second. At the end of each oculomotor status was essentially normal. presentation, the subject was required to repeat the list items back in their given order. The subjects' responses Head-Injured Controls were tape-recorded and then transcribed independently As a control group for HM, 43 nonamnesic head-injured by a research assistant experienced in the audio tran- war veterans were recruited (mean age = 42.3, SD = scription of speech. The transcripts were then marked 6.4). All had seen active service in the Korean conflict and up, firstly in order to note whether each target word had had sustained head injuries as a result. Thirty-one had been recalled correctly or not. Since the focus of interest unilateral lesions (identified by lobe as frontal, temporal, for the present investigation was item errors, subjects' parietal, or occipital). Twelve had bilateral lesions. responses were scored using a free-recall criterion with word order errors ignored. Normal Controls To allow the most straightforward comparison to HM, results will be reported for those members of each The normal control group consisted of 18 non-head- group who had the same word span as HMÐnamely 4 injured males, also veterans of the Korean conflict items. This was, in fact, the median span for each group. (mean age = 39.6, SD = 3.1). After removing subjects with greater or lesser word spans, this left 8 Korsakoff's amnesics, 27 head-injured Span Assessment controls, and 8 normal controls. Since all subjects had a Word span was measured using the following proce- word span of 4, the span + 1 list for these subjects (List dure. Subjects were presented auditorily with blocks of 1) consisted of 20 lists of 5 words. The span + 2 list (List five lists recorded on tape, starting with word lists of 2) consisted of 20 lists of 6 words. The full set of words length two and increasing the number of words in the for each list is shown in the Appendix. Characteristics of list by one at each new block. The instructions were to the words usedÐmean length in phonemes and mean listen to the taped list and then to repeat back as many frequencies in occurrences per millionÐare summarized of the words in the list as possible following their in Table 7. original order. Span was taken as the maximum list Lists 1 and 2 did not differ significantly in terms of length at which a subject managed to recall at least mean phoneme length, t(216.5) = 0.65, p > .05. List 2, 80% of lists (i.e., 4/5 lists) correctly. which contained the longer six-word sequences, con- If a subject managed to recall 80% or more of the lists tained words that had a slightly higher mean frequency, correctly another block of five was administered at the t(168.3) = 2.04, p < .05. next list length. It should be noted that taking span as the maximum list length at which a subject manages to Table 7. Phoneme and Frequency Characteristics of Words recall 80% of lists provides a slightly more conservative Used in Serial Recall Lists measure of span than the 50% criterion, which is Number of Phonemes Frequencya frequently used in other studies. Using this strict crite- rion, HM's word span was assessed as 4. Mean SD Range Mean SD Range

The Supraspan Serial Recall Task List 1 3.42 0.68 2±5 59.8 70.9 0±500 Materials List 2 3.48 0.76 2±6 93.6 164 0±1207

The words used in the immediate serial recall task were aFrequencies in occurrences per million from Kucera and Francis selected from a pool of the 800 most frequent nouns (1967).

Knott and Marslen-Wilson 605

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 APPENDIX

List 1 List 2 boy lace judge steer type plate truck bond aim wealth case oil pair gum rain code fame scar pint brush clock hate flag priest aid scare glass shock grain ball suit lie pest tire lamp fetch plant heel boat fold tax dress height march Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 stove mask gift force crew land fine storm birth crime heart chance bomb score fist tool ground wrench date joke sail mole rib sort farm inch nurse patch bend course mind year toast dust cheese twin sight rag couch tongue list sand fuse worth claim task hill fund walk nerve bid shake trial wife crash month chip yell dawn lip sack log coast gleam bill mist song faint term pitch wax ship web life tramp cave roll trade skill fence word pant edge fruit play latch dime style coach mark shine price thread hen ace stroke turn pledge man part seed rank mile tooth job silk wheel camp drink bath mail sheet grass desk bone night hook chief size marsh glue sink cure trick wing yard splash deed bribe map coin throat base pile mouse deal flood frost kid scale bowl art mass league snow joy mat pinch block choice goat hand vase load vest rock catch space broom treat germ park lad share field key bell chest sleep wind dock fish school kiss branch print

Acknowledgments 2. On this account phonological, lexical, and semantic levels of representation are each assumed to have some degree The research reported here with the patient HM was carried of storage capacity by virtue of their ability to sustain a out over the period 1970±1973 while the second author was a representation in a heightened state of activation for a short graduate student in the Department of Psychology at MIT. The period of time. research was carried out in collaboration with the late 3. Similar errors may also be observed in unimpaired Professor Hans-Lukas Teuber, and with the invaluable support subjects when the capacity of short-term memory becomes and advice of Dr. Sue Corkin. We are also greatly indebted to sufficiently overloaded. While this may occasionally occur for Ms. Susan Stackland, who carried out the testing of the head- span or even subspan lists, the frequency of such errors is injured and non-head-injured control subjects at the Clinical expected to be low. Research Center at MIT, and who also transcribed the 4. It is important to note that these phonological errors audiotapes of all of the participants' responses. We thank the occur in the complete absence of similar errors in either staff of the CRC for their help throughout, with the support of spontaneous speech or naming. NIH Grant RR-00088. We also acknowledge the help of Dr. 5. Cermak (1982) and Parkinson (1982) both note that, Nelson Butters in making available, through the VA system at although within the normal range, amnesic digit spans often Boston and Brockton, the Korsakoff patients tested during this fall at the low end of the normal distribution. This period, and we thank Professor Brenda Milner for giving us observation is consistent with current views of immediate access to HM. We also thank HM and his mother, for their serial recall that suggest a small but measurable long-term courtesy and patience during the visits to test HM in his home. memory contribution even for span length lists (Hulme et al., Reprint requests should be sent to William Marslen-Wilson, 1991, 1997). MRC Cognition and Brain Sciences Unit, 15 Chaucer Road, 6. Two-way ANOVAs comparing the two control groups Cambridge CB2 2EF, UK. E-mail: william.marslen-wilson@ (normal and head-injured) revealed no significant difference in mrc-cbu.cam.ac.uk. their scores at either list length, List 1: F(1,33) = 0.97, MSE = .037, p > .05; List 2: F(1,33) = .013, MSE = .06, p > .05. Although a strong serial position effect was apparent, there was Notes also no interaction between serial position and control group, List 1: F(1,33) = 0.83, MSE = .009, p > .05; List 2: F(1,33) = 1. Transcription symbols are based on those used in Wells .01, MSE = .0004, p > .05. In view of the absence of any (1990). important difference between the two control groups, in

606 Journal of Cognitive Neuroscience Volume 13, Number 5

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 subsequent ANOVAs they will be treated as a single group. the medial temporal lobe: A simple network model. Pro- Overall means are reported where appropriate. ceedings of the National Academy of Sciences, U.S.A., 91, 7. Imageability values are ratings on a scale of 100±700 7041±7045. taken from the Oxford Psycholinguistic Database (Quinlan, Baddeley, A. D., & Hitch, G. (1974). Working memory. In G. A. 1992). Statistics are based on the subset of items for which Bower (Ed.), Recent advances in learning and motivation imageability values were available (List 1, n = 80; List 2, n = 103). (vol. 8, pp. 47±90). New York: Academic Press. 8. Although we have not adopted a syllabic position Baddeley, A. D., & Warrington, E. K. (1970). Amnesia and the constraint in the marking of retained phonemes, we should distinction between long-and short-termmemory. Journal note that some models, such as Mackay et al.'s (1998a, 1998b) of Verbal Learning and Verbal Behavior, 9, 176±189. NST make predictions concerning phoneme position coding. Bourassa, D. C., & Besner, D. (1994). Beyond the articu- In this case, ``dog'' recalled as ``lard'' may share no phonemes if latory loop: A semantic contribution to serial order recall word initial ``d'' differs from word final ``d.'' The errors of of subspan lists. Psychonomic Bulletin and Review, 1, Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 aphasic speakers may not, however, always adhere to these 122±125. syllable position constraints, and so we have avoided such a Burgess, N., & Hitch, G. J. (1992). Towards a network model of strict criterion when marking retained phonemes. Given the the articulatory loop. Journal of Memory and Language, 31, general preservation of syllable position information in HM's 429±460. errors, this choice is unlikely to have any significant bearing on Butters, N. (1984). Alcoholic Korsakoff's syndrome: An update. the results. Seminars in Neurology, 4, 229±247. 9. HM made 3 errors out of 48 responses; these were all Cermak, L. S. (1982). The long and short of it in amnesia. failures to respond to a target. In L. S. Cermak (Ed.), Human memory and amnesia 10. HM's phonological errors were frequently characterized (pp. 43±56). Hillsdale, NJ: Erlbaum. by blending of the phonemic constituents of separate list Cermak, L. S., & Butters, N. (1973). Information processing targets (e.g., ``snow joy'' ! /dkoP/). The susceptibility of deficits of alcoholic Korsakoff patients. Quarterly Journal of amnesic patients to errors that blend components of separate Studies on Alcohol, 34, 1110±1132. stimuli has previously been noted in a recognition memory Corkin, S. (1984a). H.M.'s detection and description of am- task described by Kroll, Knight, Metcalfe, and Wolf (1996). biguous meanings. Unpublished transcript, Department of Compared to controls or patients with right hemisphere Brain and Cognitive Sciences, MIT, Cambridge. lesions, patients with left hippocampal lesions were much Corkin, S. (1984b). Lasting consequences of bilateral medial more likely to classify as ``old,'' previously unseen verbal temporal lobectomy: Clinical course and experimental find- conjunctions where the conjunction was formed by blending ings in H.M. Seminars in Neurology, 4, 249±259. constituents of previously seen words (e.g., ``valley'' + Corkin, S., Amaral, D. G., Gilberto Gonzalez, R., Johnson, K. A., ``barter'' producing ``barley''). A further important difference & Hyman, B. T. (1997). H.M.'s medial temporal lobe lesion: from the present study was that these conjunction errors were Findings from magnetic resonance imaging. Journal of observed at significantly longer timescales. Thus, the recogni- Neuroscience, 17, 3964±3979. tion probe was presented between 5 and 40 words (approxi- Dell, G. S., & O'Seaghda, P. G. (1992). Stages of lexical access in mately 9±70 sec) after the second constituent of the language production. Cognition, 42, 287±314. conjunction. There is nevertheless an intriguing similarity Drachman, D. A., & Arbit, J. (1966). Memory and the hippo- between the blend errors observed in HM's supraspan serial campal complex: II. Is memory a multiple process? Archives recall and these recognition memory errors. Interestingly, Kroll of Neurology, 15, 52±61. et al. also attribute conjunction errors in recognition memory Ellis, N. C., & Sinclair, S. G. (1996). Working memory in the to a defective consolidation process. Notably, however, where acquisition of vocabulary and syntax: Putting language in we assume phonological errors in HM's serial recall arise from good order. Quarterly Journal of Experimental Psychology, inadequate consolidation, Kroll et al. see conjunction errors in 49A, 234±250. recognition memory as a result of disinhibition of consolida- Gabrieli, J. D. E., Cohen, N. J., & Corkin, S. (1988). The im- tion, which leads to erroneous consolidations. paired learning of semantic knowledge following medial 11. The anatomical basis of Korsakoff's amnesia remains a temporal lobe resection. Brain and Cognition, 7, 157±177. matter of some debate. Adams and Victor (1993) attributed Gathercole, S. E., & Adams, A. M. (1994). Children's phonolo- Korsakoff's amnesia to lesions of the medial thalami, while gical working memory: Contributions of long-term knowl- other studies have proposed parallel damage to thalamic edge and rehearsal. Journal of Memory and Language, 33, structures and the mammillary bodies (Butters, 1984; Mishkin, 672±688. 1982). Many authors have argued for additional cortical, Gathercole, S. E., & Baddeley, A. D. (1989). Evaluation of the particularly frontal lesions, but it seems unlikely that these role of phonological STM in the development of vocabulary additional impairments bear an important relationship to the in children: A longitudinal study. Journal of Memory and patients' anterograde amnesia (Paller et al., 1997). Meanwhile, Language, 28, 200±213. a small number of studies have also found evidence suggestive Gathercole, S. E., & Baddeley, A. D. (1990). The role of pho- of hippocampal involvement (Jernigan, Schafer, Butters, & nological memory in vocabulary acquisition: A study of Cermak, 1991; Victor, Adams, & Collins, 1989; Mayes, Meudell, young children learning arbitrary names of toys. British Mann, & Pickering, 1988). More recent functional imaging Journal of Psychology, 81, 439±454. studies have, however, failed to find evidence for general Gathercole, S. E., & Baddeley, A. D. (1993). Working memory hippocampal involvement (Paller et al., 1997). and language. Hove: Erlbaum. 12. S. Corkin (personal communication). Gathercole, S. E., Service, E., Hitch, G. J., Adams, A, & Martin, A. J. (1999). Phonological short-term memory and vocabu- lary development: Further evidence on the nature of the REFERENCES relationship. Applied Cognitive Psychology, 13, 65±77. Gathercole, S. E., Willis, C., Emslie, H., & Baddeley, A. (1992). Adams, R. D., & Victor, M. (1993). Principles of neurology (5th Phonological memory and vocabulary development during ed.). New York: McGraw-Hill. the early school years: A longitudinal study. Developmental Alvarez, P., & Squire, L. R. (1994). Memory consolidation and Psychology, 28, 887±898.

Knott and Marslen-Wilson 607

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 Hulme, C., Maughan, S., & Brown, G. D. A. (1991). Memory for mon underlying processes. Cognitive Neuropsychology, 14, familiar and unfamiliar words: Evidence for a long-term 641±682. memory contribution to short-term memory span, Journal Martin, R. C., & Lesch, M. F. (1996). Associations and disso- of Memory and Language, 30, 685±701. ciations between language impairment and list recall: Im- Hulme, C., Roodenrys, S., Schweikert, R., Brown, G. D. A., plications for models of STM. In S. E. Gathercole (Ed.), Martin, S., & Stuart, G. (1997). Word-frequency effects on Models of short-term memory. (pp. 149±178) Hove, UK: short-term memory tasks: Evidence for a redintegration Psychology Press. process in immediate serial recall. Journal of Experimen- Mayes, A. R., Meudell, P. R., Mann, D., & Pickering, A. (1988). tal Psychology, Learning, Memory and Cognition, 23, Location of lesions in Korsakoff's syndrome: Neuropsycho- 1217±1232. logical and neuropathological data on two patients. Cortex, Jernigan, T. L., Schafer, K., Butters, N., & Cermak, L. S. (1991). 24, 367±388. Magnetic resonance imaging of alcoholic Korsakoff patients. McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021 Neuropsychopharmacology, 4, 175±186. Why there are complementary learning systems in the hip- Kitchener, E. G., Hodges, J. R., & McCarthy, R. (1998). Acqui- pocampus and neocortex: Insights from the successes and sition of post-morbid vocabulary and semantic facts in the failures of connectionist models of learning and memory. absence of episodic memory. Brain, 121, 1313±1327. Psychological Review, 102, 419±457. Knott, R., Patterson, K., & Hodges, J. R. (1997). Lexical and Milner, B., Corkin, S., & Teuber, H. L. (1968). Further analysis semantic binding effects in short-term memory: Evidence of the hippocampal amnesic syndrome: 14-year follow-up from semantic dementia. Cognitive Neuropsychology, 14, study of H.M. Neuropsychologia, 6, 215±234. 1165±1216. Mishkin, M. (1982). A memory system in the monkey. Philo- Knott, R., Patterson, K., & Hodges, J. R. (1999). Words in sophical Transactions of the Royal Society, London, B298, phonological memory: What preserves their phonological 85±95. structure? Unpublished manuscript. Nadel, L., & Moscovitch, M. (1997). Memory consolidation, Knott, R., Patterson, K., & Hodges, J. R. (2000). The role of retrograde amnesia and the hippocampal complex. Current speech production in auditory±verbal short-term memory: Opinion in Neurobiology, 7, 217±227. Evidence from progressive fluent aphasia. Neuropsycholo- Paller, K. A., Acharya, A., Richardson, B. C., Plaisant, O., gia, 38, 125±142. Shimamura, A. P., Reed, B. R., & Jagust, W. J. (1997). Kroll, N. E. A., Knight, R. T., Metcalfe, J., & Wolf, E. S. (1996). Functional neuroimaging of cortical dysfunction in alcoholic Cohesion failure as a source of memory illusions. Journal of Korsakoff's syndrome. Journal of Cognitive Neuroscience, Memory and Language, 35, 176±196. 9, 277±293. Kucera, H., & Francis, W. N. (1967). Computational analysis of Papagno, C., & Vallar, G. (1992). Phonological short-term present-day American English. Providence, RI: Brown Uni- memory and the learning of novel wordsÐthe effect of versity Press. phonological similarity and item length. Quarterly Journal Lackner, J. R. (1974). Observations on the speech processing of Experimental Psychology, 44A, 47±67. capabilities of an amnesic patient: Several aspects of HM's Parkinson, S. R. (1982). Performance deficits in short-term language function. Neuropsychologia, 12, 199±207. memory tasks: A comparison of amnesic Korsakoff patients Leng, R. C., & Parkin, A. J. (1989). Aetiological variation in the and the aged. In L. S. Cermak (Ed.), Human memory and amnesic syndrome: Comparisons using the Brown±Peterson amnesia (pp. 77±94). Hillsdale, NJ: Erlbaum. task. Cortex, 25, 251±259. Patterson, K., Graham N., Hodges, J. R. (1994). The impact of Mackay, D. G. (1972). The search for ambiguity by an semantic memory loss on phonological representations. amnesic patient: Implications for the theory of compre- Journal of Cognitive Neuroscience, 6, 57±69. hension, memory and attention. Unpublished manu- Poirier, M., & St. Aubin, J. (1995). Memory for related and script, Psychology Department, University of California, unrelated words: Further evidence on the influence of se- Los Angeles. mantic factors in immediate serial recall. Quarterly Journal Mackay, D. G. (1987). The organization of perception and of Experimental Psychology, 48A, 384±404. action: A theory of language and other cognitive skills. New Poirier, M., & St. Aubin, J. (1996). Immediate serial recall, word York: Springer-Verlag. frequency, item identity and item position. Canadian Mackay, D. G., Burke, D. M., & Stewart, R. (1998a). H.M.'s Journal of Experimental Psychology [Revue Canadienne de language production deficits: Implications for relations be- Psychologie Experimentale], 50, 408±412. tween memory, semantic binding and the hippocampal Postle, B. R., & Corkin, S. (1998). Impaired word-stem com- system. Journal of Memory and Language, 38, 28±69. pletion priming but intact perceptual identification priming Mackay, D. G., Stewart, R., & Burke, D. M. (1998b). H.M. re- with novel words: Evidence from the amnesic patient H.M. visited: Relations between language comprehension, mem- Neuropsychologia, 36, 421±440. ory and the hippocampal system. Journal of Cognitive Quinlan, P. (1992). The Oxford Psycholinguistic Database. Ox- Neuroscience, 10, 377±394. ford: Oxford University Press. Marslen-Wilson, W. (1970). Biographical interviews with H.M Saffran, E., & Martin N. (1990). Neuropsychological evidence for (pp. 1±144). Unpublished transcript. lexical involvement in STM. In G. Vallar & T. Shallice (Eds.), Marslen-Wilson, W., & Teuber, H.-L. (1975). Memory for re- Neuropsychological impairments of short-term memory mote events in anterograde amnesia: Recognition of public (pp. 145±166). Cambridge: Cambridge University Press. figures from news photographs. Neurospychologia, 13, Scoville, W. B., & Milner, B. (1957). Loss of recent memory after 353±364. bilateral hippocampal lesions. Journal of Neurology, Neu- Marslen-Wilson, W. D., & Tyler, L. K. (1975). Processing struc- rosurgery, and Psychiatry, 20, 11±21. ture of sentence perception. Nature, 257, 784±786. Squire, L. R. (1992). Memory and the hippocampus: A synthesis Marslen-Wilson, W. D., & Tyler, L. K. (1980). The temporal from findings with rats, monkeys and humans. Psychological structure of spoken language understanding. Cognition, 8, Review, 2, 195±231. 1±71. Thorndike, E. L., & Lorge, I. (1944). The teacher's word book Martin, N., & Saffran, E. M. (1997). Language and auditory± of 30,000 words. New York: Teachers College, Columbia verbal short-term memory impairments: Evidence for com- University.

608 Journal of Cognitive Neuroscience Volume 13, Number 5

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021 Treiman, R., & Danis C. (1988). Short-term memory errors for orders due to alcoholism and malnutrition (2nd ed.) spoken syllables are affected by the linguistic structure of the Philadelphia: F.A. Davis. syllables. Journal of Experimental Psychology, Learning, Warrington, E. K. (1982). The double dissociation of short-and Memory, and Cognition, 14, 145±152. long-term memory deficits. In L. S. Cermak (Ed.), Human Verfaellie, M., Croce, P, & Milberg, W. P. (1995). The role of memory and amnesia (pp. 61±76). Hillsdale, NJ: Erlbaum. episodic memory in semantic learning: An examination of Wells, J. C. (1990). The Longman pronunciation dictionary. vocabulary acquisition in a patient with amnesia due to en- UK: Longman. cephalitis. Neurocase, 1, 291±304. Wickelgren, W. A. (1968). Sparing of short-term memory in an Victor, M., Adams, R. D., & Collins, G. H. (1989). The Wer- amnesic patient: Implications for the strength theory of nicke±Korsakoff syndrome and related neurological dis- memory. Neuropsychologia, 6, 235±244. Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/13/5/593/1760383/089892901750363181.pdf by guest on 18 May 2021

Knott and Marslen-Wilson 609

Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/089892901750363181 by guest on 30 September 2021