Hippocampus supports long-term maintenance of language representations: Evidence of impaired collocation knowledge in amnesia

Abstract

Traditional systems consolidation theories of memory suggest that the role of the hippocampus in maintaining memory representations diminishes over time, with learned information eventually becoming fully independent of the hippocampus. Knowledge of collocations in one’s native (L1) language are acquired during development and are solidly acquired by adulthood. Remote semantic knowledge of collocations might therefore be expected to be resistant to hippocampal pathology. Patients with hippocampal damage and severe anterograde amnesia completed two tasks testing English collocation knowledge originally designed for use with English language learners. Patients with hippocampal damage demonstrated impairments in recognition of common English collocations, despite a lifetime of language experience (including postsecondary education) prior to sustaining this damage. These results suggest the hippocampus contributes to the long-term maintenance of linguistic representations and provides a challenge to traditional consolidation views of memory and an extension of newer theories to include a role for the hippocampus in maintaining semantic memory.

Graphical Abstract

Language and memory have historically been studied apart as distinct cognitive abilities with distinct neural substrates, and with distinct research methods. Over the past several decades, however, a growing body of evidence suggests that language and memory are more interdependent and intertwined than previously thought, can be impaired in tandem, and may even rely on shared cognitive and neural mechanisms. Patient studies have figured prominently in advances characterizing the language-memory interface, particularly patients with hippocampal dysfunction and deficits in relational memory binding (Banjac et al., 2021; Duff & Brown-Schmidt, 2012; Duff & Covington, in press; MacKay et al, 1998; Piai et al., 2016; Roger et al., 2022). Multiple threads of the seminal work of Sergio Della Sala are evident in this emerging literature including the value of studying neurological patients and cases of amnesia to gain novel insights about cognition (e.g., MacPherson & Della Sala, 2019), the role of representational binding deficits in mechanistic accounts of behavioral dysfunction (e.g., Della Sala et al., 2012), and what forgetting tells us about the fate of memory representations between initial acquisition and later remembering (e.g., Della Sala, 2010; Della Sala et al., 2024).

Della Sala’s work around what it means to forget - to have previously acquired information not be available, wholly or in part, either temporarily or permanently (Della Sala & Cubelli, 2021) - is particularly relevant to understanding aspects of the interaction of language and memory. There are many examples of everyday failures of memory across contexts of language use such as not recalling the name of an object in a conversation or the phenomena of language attrition observed in some multilingual speakers. There are also catastrophic failures of memory, or amnesia, associated with brain injury or disease that can prevent individuals from acquiring new language representations and can strip them of previously acquired knowledge (Duff & Cohen, 2020). In the current study, we bring these threads together to examine the status of collocation knowledge, i.e., what we know about the combinations of words that can co-occur with one another in our language, in individuals who have been living with hippocampal amnesia for over 20 years. For example, in English take precautions is judged by a majority of speakers as an acceptable collocation; likewise, a majority of speakers would reject the phrase shed attention as not acceptable. While initial collocation knowledge is built up over the course of development, this knowledge is maintained, refined, and reinforced by decades of language exposure and experience. Examining collocation knowledge in amnesia provides a unique opportunity to assess a role for the hippocampus in refining and maintaining relational knowledge of language representations beyond their initial acquisition.

Extending the Reach of the Hippocampus and Hippocampal Functionality to Language

The critical role of the hippocampus (and related medial temporal lobe structures) in the formation, and subsequent retrieval, of new enduring long-term memory is well established. The hippocampus plays a central role in support of relational memory binding (Cohen & Eichenbaum, 1993; Cohen et al., 1997; Davachi, 2006; Rubin & Cohen, 2017; Yonelinas et al., 2019), which permits the encoding of the co-occurrences of people, places, and things, along with the spatial, temporal, and interactional relations among them, into long-term declarative (episodic) memories (Konkel et al., 2008). The hippocampus supports the rapid acquisition, flexible expression, and conscious awareness of these relational memory representations providing the basis for the larger record of one’s experience in the world (Cohen & Eichenbaum, 1993; Eichenbaum & Cohen, 2001; Reber et al., 1996; Squire & Zola, 1996).

From this historical perspective on the functionality of the hippocampus above, it is counterintuitive to predict any role for it in collocation knowledge, either in its initial acquisition or its long-term tuning and maintenance. Collocations are combinations of words that co-occur with one another over and above what would be expected based on the frequencies of their constituent parts (Hoey, 1991; Jones & Sinclair 1974; Kjellmer 1990). This knowledge is built up during development and then maintained through extensive language experience and exposure across the lifespan. Native (or “L1”) speakers of a language tend to recognize common collocations as “acceptable” at high levels of performance (Gyllstad, 2007; Yamashita & Jiang, 2010). For example, L1 speakers of English will endorse the combination of words in the phrase make a mistake as acceptable and reject the phrase do a mistake as not acceptable. Language users slowly and incrementally acquire, refine, and maintain knowledge about semantic and syntactic relations to make these judgements based on their frequencies of use and the statistical regularities of the language. Critically, most language users do not have conscious or explicit knowledge of when this information was acquired or why some collocations are acceptable and others are not.

Rather, descriptions of collocation knowledge fit well with statistical learning processes which are critical to the acquisition, use, and maintenance of language representations as individuals extract the underlying patterns of linguistic co-occurrence from their frequencies in the ambient language (e.g., Saffran et al., 1996; Conway & Christiansen, 2005, Seidenberg & MacDonald, 2018; Yi, 2018). Statistical learning has been linked to other implicit learning paradigms, such as artificial grammar learning, and to the neural substrates of the basal ganglia (Evans et al., 2009; Kim et al., 2009; Perruchet & Pacton, 2006). The long-term maintenance and tuning of statistical frequencies would also seem to require a memory system that can contribute to co-occurrence knowledge in the moment and over time to update knowledge with additional experience. Thus, the demands of acquiring and maintaining collocation knowledge would seem to best fit with the incremental and implicit processes associated with the non-declarative memory system (Knowlton & Squire, 1996; Reber et al., 1996).

So why investigate the impact of hippocampal amnesia on the integrity of collocation knowledge? Over the past several decades, research has revealed the extended reach of the hippocampus and the expanded breadth of hippocampal functionality in a surprisingly broad way across cognitive domains (see Hannula & Duff, 2017; Moscovitch et al., 2016; Rubin et al., 2014; Rubin & Cohen, 2017 for reviews). We have been particularly interested in the implications of this emerging view of the hippocampus for understanding the neural substrates of language representation, use, and processing (Duff & Covington, in press; Duff & Brown-Schmidt, 2012; 2017). There are several related discoveries that motivate a role for the hippocampus in collocation knowledge and a prediction that hippocampal damage may disrupt it.

First, research has challenged the idea that the province of the hippocampus is exclusively conscious, or explicit, long-term memory showing that it also plays a critical role in memory for the unconscious, or implicit, processing of relational binding and in memory for relations over very short delays (or even no delays at all, on the timescale of short-term or working memory) (see Hannula & Green, 2012; Hannula et al., 2017 for reviews). Of particular interest is evidence linking statistical learning of co-occurrence information to a neural network that includes the hippocampus (e.g., Turk-Browne et al., 2009) and work demonstrating that hippocampal amnesia disrupts statistical learning (Covington et al, 2018; Schapiro et al., 2014).

Second, recent work has challenged the traditional view of the hippocampus in memory consolidation which posited a critical but time limited role of the hippocampus in the initial acquisition of new memories after which these memories become independent of the hippocampus over time through neocortical consolidation processes (McClelland et al., 1995). There is now evidence that the hippocampus is involved in the long-term, continuous, maintenance and retrieval of hippocampal-dependent memory (e.g., Nadel & Moscovitch, 1997; Winocur et al., 2010; Yonelinas et al., 2019). This long-term involvement fits with work showing hippocampal involvement in the updating and maintenance of relational information in the moment (Hannula et al., 2006, Warren et al., 2011) and to the updating and strengthening of previously acquired information through reconsolidation over time (McKenzie & Eichenbaum, 2011, Lee, 2008). Furthermore, the hippocampus is required for normal memory consolidation even when it was not required for successful learning during initial acquisition (e.g., a sequence learning task; Schapiro et al., 2019).

Taken together, research linking the hippocampus to implicit learning of co-occurrence information and the long-term updating and maintenance of relational knowledge over time raises the possibility that the hippocampus may play a role in the maintenance and reinforcement of knowledge about linguistic collocates. If so, hippocampal damage may cause disruptions (i.e., weakening, attrition, forgetting) to collocation knowledge acquired long before the onset of amnesia.

The Role of the Hippocampus in Language Representations Over Time

While most of the work demonstrating long-term involvement of the hippocampus to relational information comes from studies of episodic memory (autobiographical knowledge of the events of one’s life), there is emerging evidence that the hippocampus may also support semantic memory representations (knowledge of words, concepts, and general facts) over time. Individuals with amnesia recount significantly fewer details from highly familiar fairy tales and bible stories than controls (Ronsenbuam et al., 2009; Verfaellie et al., 2014). Individuals with hippocampal amnesia also produce significantly less semantic information associated with highly familiar words than do comparison participants. In one study, we assessed remote semantic knowledge by asking individuals with hippocampal amnesia to complete three tasks: a word features task (name all of the features of a word; e.g., “lemon” tastes sour, is native to Asia, used in tea); a word senses task (name all the senses of a word; e.g., lemon can be a fruit, a color, a defective automobile) and a word associates test (identifying synonyms and acceptable collocates for a target word) (Klooster & Duff, 2015).

Participants with hippocampal amnesia performed significantly worse than both neurotypical and brain-damaged comparison groups, on all three measures of word knowledge. For example, individuals with amnesia generated, on average, only half the number of features for common words (e.g., shirt) as either comparison group. This deficit was evident despite showing no differences from comparison participants on self-reported rates of familiarity of words used in the word features and senses tasks (scoring familiarity with a particular word on a 9-point scale) suggesting that these words were first learned long before the onset of their amnesia. These findings suggested that remote semantic memory is impoverished in hippocampal amnesia and that the hippocampus may play a long-term role in the maintenance of semantic information over time (Klooster & Duff, 2015).

This observed impairment in hippocampal amnesia could be a deficit in enriching (i.e., learning new features) or in maintaining semantic knowledge (i.e., language attrition) over time. An ideal method to adjudicate these alternative interpretations would be to systematically and longitudinally probe semantic knowledge from the onset of hippocampal damage and amnesia. This type of study, for our group at least, is not possible as our participants have been living with hippocampal damage and their memory impairment for over 20 years. Instead, we recruited a new group of comparison participants. Whereas in Klooster and Duff (2015) we matched the participants with amnesia to a comparison group who were the same chronological age as the participants with hippocampal amnesia at the time of testing (Age-Matched), we next recruited comparison participants who were matched to each patient at the age of onset of amnesia (Onset-Matched; Klooster et al., 2020). This new comparison group was the same age and had the same level of education as the patients when they incurred their brain damage and amnesia and they completed the same word features, word senses, and word associates test. We thought that if the participants with hippocampal amnesia performed comparably to the Onset-Matched group it would suggest that the deficit was in acquiring new semantic features over time; that is, the hippocampus is required for updating and refining of semantic representations based on new input, but not for the maintenance of previously-learned aspects of semantic knowledge. However, if participants with hippocampal amnesia performed more poorly than the Onset-Matched group, it could suggest a loss of, or failure to maintain, previously acquired knowledge.

We found that the older Age-Matched comparison group performed significantly better than the younger Onset-Matched comparison group on all tasks suggesting that with age, neurotypical individuals add new features, senses, and associates to existing semantic representations (Klooster et al., 2020). The previously reported difference in performance between the participants with amnesia and their chronological Age-Matched comparison participants becomes all the more striking as it highlights the amount of semantic knowledge the participants with hippocampal amnesia would have likely acquired over time if not for their hippocampal pathology. However, we found that the Onset-Matched group of comparison participants, matched to the age of patients at the time of onset of hippocampal damage and amnesia, also performed significantly better than the participants with hippocampal amnesia. This finding raises the possibility that hippocampal damage is associated with language loss or attrition. That is, despite having 20 additional years of language experience, individuals with amnesia presented with significantly less language knowledge than the Onset-Matched comparison participants.

The Current Study

Here, we examine whether hippocampal damage in adulthood negatively impacts recognition of English collocations. Following Klooster and colleagues (2020), we compare the performance of patients who have been living with hippocampal damage and severe anterograde amnesia for over twenty years to two neurotypical comparison groups: individuals who were the same chronological age as the participants with amnesia at the time of testing (Age-Matched) and individuals who were the same age as the patients at the time they acquired hippocampal damage (Onset-Matched). We hypothesized that the hippocampus plays a critical role in the long-term maintenance and refining of knowledge about the co-occurrence of linguistic information over time and thus, predicted that hippocampal damage will impair the recognition of collocations in individuals with hippocampal amnesia.

This prediction is supported, in part, by findings of impairment in amnesia on the Word Associates Task reported in Klooster and Duff (2015). The nature of that task, however, makes it difficult to evaluate the specific contribution of the hippocampus to collocation knowledge. In the Word Associates Task, knowledge of synonyms and collocations are assessed together and performance is commonly reported in a single composite score (Read 1993; 1998). For example, for each item, a target adjective is presented (e.g., sudden) and below it are eight possible associates; four possible synonyms to the target (e.g., beautiful, quick, surprising, thirst) and four possible collocates (change, doctor, noise, school). In contrast, the current study uses tasks focused specifically on English collocation knowledge that were originally developed for assessment of English language learners (Gyllstad, 2007). The collocations under study here (verb + noun collocates) also differ from those incorporated into the Word Associates Task (adjective + noun collocates). Use of the Gyllstad materials allows for a specific investigation of hippocampal contributions to collocation knowledge and allows us to compare the knowledge of collocations in amnesia with the data we collected from our two comparison groups as well as with learners of English with different levels of English exposure and proficiency from the original Gyllstad study.

Methods

Participants

Participants were a convenience sample of three (one female, two males) well-characterized individuals with bilateral hippocampal damage and severe anterograde memory impairment, or amnesia. At the time of data collection, the participants with hippocampal amnesia ranged in age from 53 to 64 years old (see Table 1). The age of onset of hippocampal damage and resulting amnesia for participants 1846, 1951, and 2563 (in years) were 30, 28, and 45, respectively. Thus, at the time of the study, patients 1846, 1951, and 2563 had lived with anterograde amnesia for 23, 36, and 16 years, respectively. Prior to the onset of their amnesia, all participants had completed post-secondary education. Patient 1846 completed a two-year associate degree and patients 1951 and 2563 both completed a four-year bachelor degree.

Table 1.

F = female. M = male. Ed = years of completed education. HSE = Herpes Simplex Encephalitis. HC = Hippocampus. MTL = Medial Temporal Lobe. HC Vol = Hippocampal Volume. Volumetric data are z-scores as measured through high resolution volumetric MRI and compared to a matched healthy comparison group (see Allen et al., 2006). N/A = no available data. WMS-III GMI = Wechsler Memory Scale-III General Memory Index. WAIS-III = Wechsler Adult Intelligence Scale -III. FSIQ = Full Scale Intelligence Quotient. VIQ = Verbal Intelligence Quotient. Vocab = Vocabulary subtest. Info = Information subtest. Bolded scores are defined as 2 or more standard deviations below normative data.

Participant	Sex	Current Age	Onset Age	Ed	Etiology	Damage	HC Vol.	WMS III GMI	WAIS III FSIQ	WAIS III VIQ	WAIS III Vocab	WAIS III Info
1846	F	53	30	14	Anoxia	Bilateral HC	−4.23	57	84	88	8	8
1951	M	64	28	16	HSE	Bilateral HC + MTL	−8.10	57	106	107	10	11
2563	M	61	45	16	Anoxia	Bilateral HC	N/A	75	102	91	9	12
AMN	1F 2M	59.33	34.33	15.33				63.00	97.33	95.33	9.00	10.33
Age-Matched Comparison participants	3F 6M	60.44		15.56
Onset-Matched Comparison Participants	2F 6M	35.75		1.550

Etiologies included anoxia/hypoxia (n=2) resulting in bilateral hippocampal damage, and herpes simplex encephalitis (HSE; n=1), resulting in more extensive bilateral medial temporal lobe damage affecting the hippocampus, amygdala, and surrounding cortices. High-resolution MRI analyses were conducted for the entire brain on 2 of the 3 patients (Allen et al., 2006; Buchanan et al., 2005). These analyses showed hippocampal volumes significantly decreased for each patient, with the studentized residual differences in hippocampal volume relative to a matched comparison group reduced by 4.23 and 8.10 z-scores for 1846 and 1951, respectively. For patient 1846, additional in-depth neuroanatomical analyses were available and other medial temporal lobe structures were judged to be within the normal range through volumetric analyses and there was no visible damage to the lateral temporal or anterior temporal lobes (Warren et al., 2012). Participant 2563 wears a pacemaker and could not undergo MRI examination. Additional neuroanatomical analyses of patient 1951 revealed complete loss of the right temporal pole and right temporal lobe whereas damage to the left temporal pole was confined to the medial polar cortex (Feinstein et al., 2010).

Performance on tests of neuropsychological functioning indicated a severe impairment in declarative memory (M=63; Wechsler Memory Scale–III General Memory Index) while performance across other cognitive domains from the Wechsler Adult Intelligence Scale-III including measures of verbal IQ, vocabulary and semantic knowledge was within normal limits (Table 1). Further, standardized neuropsychological testing and interviews with a certified speech language pathologist confirmed that the patients do not have frank¹ language deficits such as aphasia or semantic dementia.

While amnesia can be associated with social isolation (Davidson et al., 2012; Tate, 2002), and thus the potential for language deprivation, that is not the case for the participants in this study. These participants live with family or other caretakers and do not differ from neurotypical comparison participants in the size of their social networks as corroborated by family members (Beadle et al., 2022). Each has daily exposure to multiple communication partners (e.g., 1951 and 2563 participate in group activities where they live as well as routine visits and travel with family; 1846 lives with her husband, daughter, son-in-law, and grandchildren) and opportunities to engage with and use language (e.g., 2563 reads the newspaper and novels although reports he cannot recall what he reads).

Following Klooster and colleagues (2020), we recruited two groups of comparison participants: Age-Matched and Onset-Matched. The Age-Matched comparison participants were 9 individuals (3 female, 6 male) who were matched pair-wise to the participants with amnesia based on chronological age. At the time of study, individuals with amnesia were 59.33 years old, on average (range: 53–64), and had 15.33 years of education, on average. The Age-Matched comparison participants were 60.44 years old (range: 50–70), on average, and had 15.56 years of education, on average.

The Onset-Matched comparison participants were 8 individuals (2 females, 6 males) who were matched pair-wise to the participants with amnesia based on the age of onset of their hippocampal damage and amnesia. At the time of onset, participants with amnesia were 34.33 years old, on average (range: 28–45), and had 15.33 years of education, on average. The Onset-Matched comparison participants were 35.75 years old (range: 27–50), on average, and had 15.50 years of education, on average.

Comparison participants were matched pairwise to each patient with amnesia on age and education to reduce between-group demographic variability and ensure similar within-group demographic variability. Sample size for the Age- and Onset-Matched groups was based on our previous work (Klooster et al., 2015; 2020). All comparison participants were native (L1) speakers of American English recruited from the United States. Comparison participants were recruited via community flyer postings and via social media posts. Prior to enrollment, all comparison participants completed a medical history screening designed to rule out diagnoses and medications that can interfere with cognition (e.g., neurological or psychiatric conditions, developmental or learning disorders, untreated diabetes, or sleep apnea).

We also compare performance of the patient and comparison groups recruited for the current study to data reported in the original Gyllstad study characterizing collocation knowledge in native (L1) speakers of English and English language learners of varying formal language experience (Gyllstad, 2007). As reported in Gyllstad, 2007, a total of 269 participants completed both tests of collocation knowledge. 34 participants were L1 speakers of English, recruited from the University of Wales, Swansea. 209 participants were English language learners recruited from Lund University. These participants had 9–10 years of English instruction prior to their enrollment in university and were further divided into groups based on their level (first-term, second-term, etc.) of English language study at university. Finally, 26 participants were English language learners in grade 11 recruited from a local upper-secondary school in Sweden.

Experimental Measures and Procedures

All participants completed two tasks assessing English collocation knowledge, originally developed for use in English language learners (Gyllstad, 2007). In both tasks, participants were asked to read lists of potential collocations and identify which word combinations are permissible in the English language. The COLLEX task consists of 50 items. On each item, participants read three word combinations (e.g., do a mistake, make a mistake, run a mistake). Participants were told that one of three word combinations is a ‘natural and frequent word combination in the English language’ whereas the other two are not. Participants chose the word combination they thought was the most natural and frequently occurring and circled it. The COLLMATCH task consists of 100 items. On each item, participants read a single word combination and were asked to indicate whether or not it is a word combination in the English language. If they thought it was a word combination used in the English language they circled YES and if they did not think the word combination was used in English they circled NO. Some examples of the single word combinations included: catch importance, take precautions, and shed attention. The original task development studies reported by Gyllstad demonstrate the presence of strong ceiling effects in the COLLEX task. We chose to administer both tasks in line with Gyllstad’s recommendation, which both acknowledges the limitations in COLLEX and yet encourages the combined use of COLLEX/COLLMATCH as a test battery (Gyllstad, 2007).

Participants completed both tasks on paper forms. Printed task instructions were read aloud to each participant while they followed along, followed by practice test items. Participants were encouraged to ask clarifying questions during these practice trials, and the correct answer to all practice trials was provided to participants. The printed task instructions were available to participants throughout the task, to re-reference if needed. All participants completed the COLLEX task followed by the COLLMATCH task. Legal copyright restrictions prevent public archiving of COLLEX and COLLMATCH tasks, but these task materials are available in the appendices of Gyllstad, 2007.

Data Analysis

Permutation analyses were run using the infer package in R (Couch et al., 2021) to determine whether the difference in proportion correct for each task differed significantly across groups. Permutation tests are a class of non-parametric methods that do not require an assumption of normality (Fisher 1935). Here, we performed 10,000 permutations of each dataset to obtain an empirical distribution of the test statistic under the null hypothesis. Bootstrapping was used to determine 95% confidence intervals around each parameter estimate and to obtain p-values. Data collected for the current study were also compared visually to data from L1 English speakers and English-language learners reported in Gyllstad, 2007. An additional post-hoc exploratory analysis of the COLLMATCH task data was conducted during the review process. Group differences in sensitivity to acceptable collocations and in response bias were assessed using signal detection theory measures, including sensitivity (d’) and response bias (c) (Stanislaw & Todorov, 1999), with log-linear corrections for extreme proportions (Hautus, 1995). Behavioral study data and analysis scripts are archived in a publicly accessible repository: https://osf.io/4trch/?view_only=334faabd75c54030a3494bdfc44c317e.

Results

COLLEX

Individual participant performance is illustrated in Figure 1. Performance of Age-Matched and Onset-Matched comparison participants was compared using permutation analyses which revealed no statistically significant difference between the two comparison groups (p = 1). Data from the two comparison groups are collapsed into a single comparison group for all subsequent analyses. Patients with hippocampal amnesia accurately responded to an average of 91.3% of COLLEX items (SD = 7.0), compared to comparison participants who responded accurately to 97.1% of items (SD = 2.4), a statistically significant difference (p = 0.003). Figure 2 illustrates the process and results of the permutation analysis of COLLEX data. In post-hoc exploratory analyses during the review process, performance of each amnesic patient was compared individually to the subset of their matched comparison participants, using the same permutation methods as described above. In these by-participant analyses, only patient 1846 performed significantly more poorly than her matched comparison participants (p = 0.02), while there was no statistically significant difference in performance for patients 1951 (p = 0.11) and 2563 (p = 1) compared to each of their matched comparison groups.

Figure 1. COLLEX Behavioral Performance.

Boxplots illustrate group median performance and interquartile range. Individual participant data is depicted using shapes that allow for comparison of pair-wise matching across groups. Please note that the y-axis is truncated at 66%.

Figure 2. COLLEX Permutation Analysis.

We used a nonparametric sampling technique to examine group differences in proportion correct on the COLLEX task. The observed difference in proportion correct for patients with amnesia was −5.8% which served as our test statistic (illustrated here with the solid red vertical line). We conducted a permutation test in which all observed values were pooled and resampled 10,000 times (the simulation-based null distribution, illustrated by the gray histogram). The difference in proportion was calculated between groups for each resampling. 34 of 10,000 resampled differences in proportion were more extreme than the test statistic, indicating that patients with amnesia were significantly less likely to respond accurately to COLLEX items relative to the healthy comparison group (p = 0.003). We then used a bootstrapping technique to estimate a confidence interval for the difference in proportion between groups. We resampled, with replacement, the difference in proportions between groups 10,000 times. We assumed that a confidence interval that did not include zero indicated a significant difference in proportions between groups. The estimated 95% confidence interval was [−11%, −1%] (illustrated here with vertical dotted red lines).

Interim Discussion of COLLEX Task Performance.

On a three-alternative forced choice collocation knowledge task, patients with amnesia, as a group, demonstrated significantly poorer collocation recognition compared to demographically-matched healthy comparison participants. Post-hoc analyses at the individual patient level indicated that this group difference was driven by the performance of patient 1846, who performed significantly more poorly on the task compared to her demographically-matched comparison participants. Patient 1951 performed numerically worse than his matched comparison participants, though this did not reach statistical significance, and 2563 performed numerically similar to his matched comparison participants.

What accounts for the difference in performance on the COLLEX task among patients? As noted above, one possible explanation is the observed ceiling effects of this task. Studies contributing to the original development of the COLLEX task suggest that while it produces valid and reliable scores, the strong ceiling effect may mean that the COLLEX task is “not sensitive enough to pick up subtle differences”, especially among higher-ability groups (Gyllstad, 2007). Given that both the patients and comparison participants are both lifelong L1 speakers of English, we would expect only subtle differences between the two groups. Gyllstad suggests that it is the three-alternative forced choice format of the COLLEX task that results in an easier task and thus stronger ceiling effects. Here, we chose to implement both collocation tasks because this is the first investigation of collocation knowledge in this population and in line with Gyllstad’s recommendation, which both acknowledges the limitations in COLLEX and yet encourages the combined use of COLLEX/COLLMATCH: “ COLLMATCH does not suffer from the same tendencies of ceiling effects as COLLEX, but the two tests make use of slightly different test tasks, and it is therefore recommended that they are used together in a test battery” (Gyllstad, 2007).

Another possibility is that the performance of patient 1846 is not representative of patients with hippocampal amnesia or is due to impairments beyond her hippocampal damage and severe amnesia. We do not believe that 1846 performs in a manner that is unrepresentative of patients with amnesia for the following reasons. The clinical criteria for amnesia include a disparity between the performance on a standardized measure of IQ (e.g., WAIS) and memory ability (e.g., WMS) of at least 20 points and that the memory impairment be focal (i.e., there are no impairments in other domains of cognition) (O’Conner, Verfaellie, & Cermak, 1995). All patients in our study, including 1846, meet these criteria. Her memory impairment is severe: like patient 1951 she performs more than 2.5 standard deviations below the mean. Her memory impairment is also focal: her IQ and vocabulary scores, while numerically lower than 1951 and 2563, fall within a standard deviation of the mean (“low average”) and her neuropsychological profile outside of her memory deficits is otherwise unremarkable (Warren et al., 2012). Neuroanatomically, her lesion is significantly more restricted within the medial temporal lobes, and to the hippocampus, than 1951 (Feinstein et al., 2010; Warren et al., 2012). We suspect that the performance of 1846 is related more to the sensitivity of the COLLEX task than her representativeness of patients with amnesia. Furthermore, we speculate that her performance here might seem like less of an outlier, and 2563’s performance less normal, on a task more sensitive to collocation knowledge and less vulnerable to ceiling effects.

COLLMATCH

Individual participant performance is illustrated in Figure 3. Performance of Age-Matched and Onset-Matched comparison participants was compared using permutation analyses which revealed no statistically significant difference between the two comparison groups (p = 0.09). Data from the two comparison groups are collapsed into a single comparison group for all subsequent analyses. Patients with hippocampal amnesia accurately responded to an average of 85.3% of COLLMATCH items (SD = 0.5), compared to comparison participants who responded accurately to 95.8% of items (SD = 2.2), a statistically significant difference (p < 0.001). Figure 4 illustrates the results of the permutation analysis of COLLMATCH data. In post-hoc exploratory analyses during the review process, performance of each amnesic patient was compared individually to their matched comparison participants, using the same permutation methods as described above. In these by-participant analyses, all three patients performed significantly more poorly than their matched comparison participants (all ps < 0.002).

Figure 3. COLLMATCH Behavioral Performance.

Boxplots illustrate group median performance and interquartile range. Individual participant data is depicted using shapes that allow for comparison of pair-wise matching across groups. Y-axis truncated at chance-level performance (50%).

Figure 4. COLLMATCH Permutation Analysis.

We used a nonparametric sampling technique to examine group differences in proportion correct on the COLLEX task. The observed difference in proportion correct for patients with amnesia was −10.5% which served as our test statistic (illustrated here with the solid red vertical line). We conducted a permutation test in which all observed values were pooled and resampled 10,000 times (the simulation-based null distribution, illustrated by the gray histogram). The difference in proportion was calculated between groups for each resampling. 0 of 10,000 resampled differences in proportion were more extreme than the test statistic, indicating that patients with amnesia were significantly less likely to respond accurately to COLLEX items relative to the healthy comparison group (p < 0.001). We then used a bootstrapping technique to estimate a confidence interval for the difference in proportion between groups. We resampled, with replacement, the difference in proportions between groups 10,000 times. We assumed that a confidence interval that did not include zero indicated a significant difference in proportions between groups. The estimated 95% confidence interval was [−15%, −6%] (illustrated here with vertical dotted red lines).

Given the binary nature of the COLLMATCH task, to ensure the group difference in accuracy was not driven by differences across groups in response bias (i.e. the participant’s general tendency to respond “yes” or “no”), an exploratory post-hoc analysis was conducted during the review process. Signal detection theory measures of sensitivity (d’) and response bias (c) were calculated for each participant’s response pattern (Stanislaw & Todorov, 1999). d’ values of 0 indicate chance-level performance and an inability to distinguish “signals” from “noise” (i.e. an inability to detect acceptable collocations from unacceptable word combinations), while larger d’ values indicate better performance (i.e. greater sensitivity to acceptable vs. unacceptable collocations). c is a measure of response bias in which 0 indicates that the participant favors neither the “yes” response nor the “no” response. c is measured in standard deviation units, with negative values indicating a bias toward responding “yes” and positive values of indicating a bias toward responding “no”. Patients with amnesia demonstrated significantly poorer sensitivity (d’ = 2.06) compared to comparison participants (d’ = 3.34, t(10) = 9.68, p < 0.001). In contrast, the two groups did not differ significantly in response bias (amnesic patients c = −0.10, comparison participants c = 0.11, t(5) = 1.62, p = 0.17).

Interim Discussion of COLLMATCH Task Performance.

Performance on the COLLMATCH task provides robust evidence of a deficit in collocation knowledge in individuals with amnesia. In contrast to the performance on the COLLEX task, all three patients with amnesia here performed significantly more poorly than their matched comparison participants. This finding lends support to the notion that COLLMATCH is a more sensitive measure of collocation knowledge, as reported in Gyllstad (2007). We also note that the performance of the three patients is nearly identical to one another on the COLLMATCH task supporting our speculation above that performance on the COLLEX task was more a function of task sensitivity than to differences in participant characteristics.

Performance Compared to English Language Learners

Here, we plot the performance of patients with hippocampal amnesia and the comparison participants from the current study against data reported by Gyllstad (2007). Participants in the Gyllstad study included Swedish English Language Learners (ELL) of varying experience, from grade 11 English learners to third-term undergraduate learners, as well as a group of native (L1) English speakers. The average number of accurate responses on COLLEX and COLLMATCH by group are illustrated in Figures 5 and 6. Comparison participants from our study perform comparably to L1 speakers of English from the Gyllstad study. Patients with hippocampal amnesia perform similarly to third-term undergraduate English language learners on both tasks.

Figure 5. COLLEX Comparison to Gyllstad, 2007 data.

Mean group-level performance across comparison groups from the original Gyllstad study (x-axis group labels in black) and the current study (x-axis group labels in blue). Y-axis truncated at chance-level performance (33%).

Figure 6. COLLMATCH Comparison to Gyllstad, 2007 data.

Mean group-level performance across comparison groups from the original Gyllstad study (x-axis group labels in black) and the current study (x-axis group labels in blue). Y-axis truncated at chance-level performance (50%).

Discussion

On two tasks assessing collocation knowledge, patients with hippocampal amnesia performed significantly more poorly than Age- and Onset-Matched comparison participants. On the less sensitive COLLEX task, these differences were driven by the performance of patient 1846. On the COLLMATCH task, which is less vulnerable to ceiling effects, patients with amnesia performed uniformly worse than matched comparison participants, an effect that was not driven by differences in response bias across the two groups. Deficits in collocation knowledge in the patient group are striking given that all the patients had more than twenty-five years of exposure to and use of English prior to the onset of their amnesia, including attainment of post-secondary degrees. In contrast, Age- and Onset-Matched comparison participants performed similarly to one another, and at ceiling, suggesting that collocation knowledge, as tested here, is well-established by early adulthood (e.g. age 28 and onward) in neurotypical L1 English speakers. Both comparison participant groups in our study performed similarly to L1 English speakers from the original Gyllstad study. In contrast, patients with amnesia performed most similarly to third-term English-language learners from the original Gyllstad study on COLLMATCH, despite being monolingual English speakers with a lifetime of typical English experience prior to the onset of their amnesia, and continued routine English language exposure in the years after onset. Taken together alongside the performance of Age- and Onset-Matched comparison participants, our results suggest that collocation knowledge is acquired in L1 language users by early adulthood, but that these representations must be maintained by ongoing hippocampally-mediated processes or they may become impoverished or weakened over time.

Errors made by patients with hippocampal amnesia occurred in spite of the simple nature of the task (requiring a straightforward recognition/acceptability judgment), the high frequency of the component words, and a lack of overt memory demands. As an illustrative example, “meet a need” is the 8th most frequent collocation in the COLLMATCH task, as measured using frequency data from the British National Corpus (occurring 107 times within that corpus; BNC Consortium, 2007). On this item, all of the Age- or Onset- Matched participants accurately recognized “meet a need” as an acceptable English collocation, whereas two of the patients with hippocampal amnesia inaccurately indicated that “meet a need” was an unacceptable collocation. Similarly, all of the Age- and Onset-Matched comparison participants rejected “stand an occasion” as acceptable (occurring 0 times within the BNC corpus), whereas two of the patients with hippocampal amnesia inaccurately endorsed “stand an occasion” as an acceptable word combination in English.

Our results extend prior findings suggesting that adults with hippocampal amnesia demonstrate subtle but measurable impairments in remote semantic memory (Rosenbaum et al., 2009; Verfaellie et al., 2014; Grilli & Verfaellie, 2014), and specifically in linguistic representations previously expected to be overlearned and independent of the hippocampus (Hilverman & Duff, 2021; Klooster & Duff, 2015; Klooster et al., 2020). While a limited role for the hippocampus in the initial acquisition of new vocabulary has historically been well-recognized (Gabrieli et al., 1988; Squire & Zola, 1998), its role in maintenance and updating of language representations has been obscured by the predominant use of language measures originally designed to diagnose frank¹ and unsubtle language impairments like aphasia and semantic dementia (Duff et al., 2020; Duff & Covington, in press). For example, early work demonstrated that patients with amnesia perform within normal limits on standardized neuropsychological measures of vocabulary knowledge and naming and are able to perform similarly to healthy comparison participants on simple experimental remote semantic knowledge tasks (Gabrieli et al., 1988; Kensinger et al., 2001; Manns et al., 2003; Verfaellie et al., 2000). Indeed, as noted in Table 1, participants in our study perform within normal limits on the Vocabulary subtest of the WAIS-III, in which they are asked to provide definitions for word stimuli. Results from these, and other, studies have historically led to the assumption that remote semantic knowledge acquired before the onset of hippocampal damage is intact in amnesia. Our results suggest that hippocampal damage impacts the ability of language users to continually update and maintain language representations, including the ability to recognize multi-word constructions like collocations that frequently co-occur in language input. These results extend our previous findings that patients with bilateral hippocampal damage perform significantly worse than neurotypical comparison participants on measures of semantic richness (Klooster & Duff, 2015; Klooster et al., 2020) and exhibit disruptions in naming when tested with an extensive set of stimulus items that vary on psycholinguistic variables (e.g., imagability, familiarity) compared to widely-used but truncated standardized assessments of naming (Hilverman & Duff, 2021). Additionally, the simplicity of the “yes/no” judgments in the COLLMATCH task reduce concerns that previous findings in this line of work were driven by the complexity or excessive episodic memory demands of the test phase itself.

In our prior work, we have asked whether hippocampal contributions to semantic richness exclusively support the addition of new information over the course of the lifespan (a finding which would be similar to the historically understood involvement of the hippocampus in new learning) or if hippocampal contributions are also important for the maintenance of previously-learned linguistic knowledge (such that hippocampal damage might result in attrition of previously-learned language representations) (Klooster & Duff, 2015; Klooster et al., 2020). To address this question, the inclusion of Age- and Onset-Matched comparison participants provide a window into the growing richness of language representations over time as a result of continued language exposure. In our prior work, older Age-Matched comparison participants produced more word features and senses in tasks examining semantic richness than did younger Onset-Matched participants, suggesting that continued language exposure, experience, and use increase the amount of information stored in semantic memory. Yet, in this study, patients with hippocampal amnesia performed even more poorly than Onset-Matched comparison participants, suggesting that hippocampal damage is associated not just with reduced ability to update previously-learned representations with new information but that reduced hippocampal functionality leads to attrition of linguistic representations over time (Klooster et al., 2020). As such, our results have important implications for our understanding of phenomena and theories associated with language and memory individually and for our understanding of the language-memory interface.

Neurobiological Evidence for the Role of Hippocampal-Dependent Memory in Collocation Acquisition and Maintenance

Our results provide converging neurobiological evidence that supports recent behavioral findings that hippocampal-dependent memory supports collocation knowledge (Divjak et al., 2022). Divjak and colleagues indirectly investigated the role of multiple memory systems in language processing of morphology, syntax, and collocations in L1 speakers of Polish, by testing acceptability judgments within each language domain under control and dual-task conditions. Under dual-task conditions, the authors propose that language users’ divided attention should reduce access to hippocampal-dependent declarative memory while leaving access to procedural memory unaffected. They found that collocation judgments were strongly impacted by dual-task conditions, compared to the other language domains under study, and further that strong declarative learners (categorized as such by separately administered learning and memory tasks) were more impacted by dual-task conditions than weak declarative learners. Divjak and colleagues conclude that declarative memory is critical for supporting collocation knowledge, even for highly-educated adult L1 speakers of a language, contrary to existing claims that, by adulthood, accumulated experience and language proficiency should result in independence from hippocampal-mediated representations (Divjak et al., 2022). Our results align with this proposal and provide neurobiological evidence for the necessity of the hippocampus in ongoing maintenance or strengthening of representations that support collocation knowledge in adulthood.

A series of recent studies provide evidence that (even highly familiar) semantic knowledge is maintained and adjusted in response to daily language exposure in L1 language users in adulthood (Gaskell et al., 2019; Mak et al., 2023). This work demonstrates that encountering a highly familiar word in a particular sentential context alters that word’s representation with consequences for its future usage (contextual priming). This priming effect is present shortly after this linguistic experience, and then fades over time during which the participant is awake. In contrast, when sleep immediately follows the initial exposure, modification of the word’s representation (as evidenced by priming effects) is maintained (Gaskell et al., 2019; Mak et al., 2023). The authors suggest that the role of sleep in bolstering the stability of these contextual priming effects suggests a role for the hippocampus in updating linguistic representations based on language experience. Indeed, Gaskell and colleagues write that while “previously language development might have been characterized as a steady progression towards a fairly stable state, it is not clear that such a stable state is ever achieved” (Gaskell et al., 2019).

Our results support this account of language plasticity in adulthood and suggest a role for the hippocampus in continuing to support the maintenance of linguistic representations, even after they have been well-established by years of prior experience. We suggest that hippocampal pathology impedes the updating of linguistic representations by reducing the benefits associated with long-term language exposure. On this view, language exposure and experience throughout the lifespan constitutes a stream of constant learning and updating opportunities. Patients with hippocampal pathology are deprived of these continued language learning opportunities, not due to lack of language exposure (as in language attrition observed in some multilingual speakers, see below) but as a result of neurobiological changes that disrupt relational binding and thus prevent their obtaining the benefits from typical daily exposure to language. Our results align with the contextual binding account of lexical knowledge proposed by Gaskell and colleagues, in which new experiences with language result in hippocampal-mediated promiscuous binding of words in their contexts (co-occurrence information), rather than direct updating of longer-term cortical-cortical lexical connections (Gaskell et al., 2019; Mak et al., 2023). Our results extend beyond the original arguments put forth by Gaskell and colleagues, who posit that hippocampal involvement in updating linguistic representations should be “pervasive and observed for any word as long as its surrounding contexts refine its interpretation in some way” (Mak et al., 2023; emphasis added). Our results suggest that even for standard interpretations of words-in-context (e.g. highly frequent and agreed-upon collocations), hippocampal involvement is still required for the ongoing maintenance of this co-occurrence information. That is, in healthy individuals, hippocampal-mediated linguistic representations formed during each exposure to a particular word in a particular context is subject to decay, but repeated language experiences and consolidation processes result in both updating and maintenance of rich linguistic representations. In patients with bilateral hippocampal damage, this process is disrupted, resulting in impoverished linguistic representations over time (Klooster et al., 2020).

Consonance with Evidence from Language Attrition in Multilingual Speakers

Our results also align with observed patterns of L1 language attrition in multilingual speakers who, due to a variety of life circumstances, experience reductions in L1 use and exposure in adulthood that can result in changes (either marked or subtle) to their production and comprehension of their first language compared to the normative performance of other L1 speakers (Kasparian et al., 2017; Opitz, 2016; Schmid & Jarvis, 2014; Schmid & Fägersten, 2010; Schmid & Yilmaz, 2018). Existing theories of language attrition in multilingual speakers note challenges in disentangling which decrements in L1 language performance should be attributed to reductions in exposure and experience with L1 versus changes attributed to L1/L2 interference, given that most speakers who experience significant changes in L1 exposure are also directly engaged in language learning within their new L2 environment (Gallo et al., 2021). That said, decrements in language performance have been tied to changes in language exposure and experience in dose-dependent ways (Kasparian et al.. 2017; Optiz, 2016; Schmid & Yilmaz, 2018). One particularly striking example reported by Baladzhaeva and Laufer (2018) provides strong evidence that reduced language experience alone can result in L1 attrition in monolingual speakers. In this case, Russian-speaking immigrants to Israel who did not learn Hebrew demonstrated as much L1 attrition as did Russian L1/Hebrew L2 bilingual speakers, as compared to Russian L1 speakers living in Russia. These results suggest that reductions in language exposure on their own can result in L1 language attrition (i.e. it is not a requirement that L1 speakers be influenced by their learning of an L2 to experience language attrition; instead, impoverished language exposure can be the only precipitating cause (Baladzhaeva & Laufer, 2018).

While manifestations of “language attrition” in multilingual speakers can include reductions in speed and accuracy in many aspects of language processing and production, across studies there are some commonly recognized patterns in how language attrition unfolds. In general, phonological and syntactic structures have been observed to be more resistant to attrition compared to the more prominent changes observed in L1-attriters’ use and knowledge of vocabulary, idioms, and collocations (Gallo et al.. 2021; MacWhinney, 2019). The vulnerability of some language structures (e.g. collocations) over others (e.g. syntax) does not appear to be well-explained by frequency effects alone (MacWhinney, 2019; Schmid, 2011). The Competition Model of language attrition proposed by MacWhinney suggests that collocation knowledge is supported by reciprocal connections between hippocampus and cortex, with the hippocampus’ demonstrated role in pattern separation maintaining separate patterns of connectivity for L1 and L2 collocations, even when they reside in nearby or overlapping regions of cortex (MacWhinney, 2019). These hippocampal-mediated connections prevent catastrophic interference between L1 and L2 representations, but the maintenance and strengthening of these connections relies on continued exposure to co-occurrence information. This requirement of repeated exposure and experience results in attrition of L1 collocations in cases where L1 experience and exposure is markedly reduced. Here, we empirically demonstrate such a phenomenon in monolingual speakers with hippocampal pathology. Again, rather than lack of language use or exposure resulting in attrition of a previously-acquired language, damage to the hippocampus causes a neuropathological change in relational binding and a reduced ability to benefit from typical L1 language exposure.

Implications for Theories of Hippocampal Involvement in Memory Representations over Time

There has been considerable debate regarding the extent of hippocampal involvement in memory representations over time. Whereas the standard systems consolidation theory (e.g., McClelland et al., 1995) argued for a time-limited role of the hippocampus (i.e., only during initial acquisition), there is growing recognition that the hippocampus plays a long-term role in the retrieval and reactivation of episodic memory (e.g., Nadel & Moscovitch, 1997) and is involved in the updating and strengthening of previously acquired information through reconsolidation over time (e.g., McKenzie & Eichenbaum, 2011). Most of the proposals on this time course, however, come from work on episodic memory with little substantive data from studies of semantic memory. This lack of data has significantly limited theory development and testing around the extent to which episodic and semantic memory have a shared dependence on the hippocampus over various time scales (Duff et al., 2020). For example, in their seminal paper laying out the points of similarity and divergence between standard consolidation models and their multiple trace theory, Nadel and Moscovitch (1997) noted that most studies of remote semantic knowledge do not include tests sensitive enough to detect deficits in patients with amnesia, which limits comparisons to other forms of memory. Similarly, in proposing an alternative to standard systems consolidation theory called contextual binding theory, Yonelinas and colleagues (2019) focused nearly exclusively on the role of the hippocampus in episodic memory and stated it was an open question whether their theory could be applied to semantic memory.

The findings here contribute to filling this gap by providing new empirical evidence for a long-term role of the hippocampus in maintaining remote semantic knowledge. These results extend our prior work showing a role for hippocampus in the long-term maintenance of remote lexical knowledge (e.g., the amount of information associated with a word or concept; Klooster & Duff, 2015) to include a role for hippocampus in the continuous maintenance and refining of knowledge about the co-occurrence of linguistic information over time. Hippocampal involvement in semantic memory broadly, and in collocation knowledge specifically, parallels its role in episodic memory including in relational memory binding (Eichenbaum & Cohen, 2001) whether the relational knowledge is explicit or implicit and irrespective of time constraints on processing or retrieval (Hannula & Green 2012; Hannula et al., 2017; Hannula et al., 2006). These data support a view of the hippocampus as making long-term continuous contributions to the representation of both episodic memory and semantic memory (Duff et al., 2020). We propose there is now sufficient evidence of impaired remote semantic memory in amnesia from studies of general knowledge (e.g., Blumenthal et al., 2017; Rosenbaum et al., 2009; Verfaellie et al., 2014; Grilli & Verfaellie, 2014) and linguistic knowledge (Hilverman & Duff, 2021; Klooster & Duff, 2015; Klooster et al., 2020) to generate more comprehensive theories of memory.

Limitations

One limitation of the current study is the near ceiling-level performance of the comparison participants on both tests of collocation knowledge. These ceiling effects are noted by Gyllstad for L1 speakers of English (2007). As in the original study, ceiling effects were less prominent for the COLLMATCH task compared to COLLEX (Gyllstad, 2007). These ceiling effects impede our ability to assess, as in the Klooster study, whether strengthened collocation knowledge is observable across the lifespan (Klooster et al., 2020). That is, with a more sensitive test of collocation knowledge, would the older Age-Matched comparison participants demonstrate stronger collocation knowledge compared to younger Onset-Matched comparison participants? While we are unable to test this hypothesis statistically, it is intriguing to consider the performance of our comparison participants to those in the original Gyllstad study. On the COLLMATCH task, which is less prone to ceiling effects, our older comparison participants perform numerically better than the undergraduate L1 speakers of English from the Gyllstad data set (Gyllstad, 2007). While the current study does not allow for assessment of collocation knowledge enrichment during healthy aging, our results complement those from Klooster and colleagues to demonstrate that hippocampal pathology results in attrition of remote semantic knowledge (Klooster et al., 2020).

There are certain limitations associated with patient work including the necessarily small sample size and neuroanatomical differences among. Patients with amnesia in the current study differ in their degree of medial temporal lobe damage, and we do not have detailed MRI analyses for one (2563) of the three patients. While the deficit in collocation knowledge presented here fits most parsimoniously with the patients’ shared hippocampal pathology and amnesia, future work is warranted to replicate these findings and to assess the potential contribution of disruptions in the broader cortico-hippocampal networks to deficits in collocation knowledge (Barnett et al., 2021; MacWhinney, 2019).

Another potential limitation are dialectical differences between our participants and those recruited for the Gyllstad study, given the reliance on the BNC corpus for the development of the original COLLEX and COLLMATCH test items (Gyllstad, 2007). This limitation should not significantly impact our primary research questions, in which we compared performance of patients with amnesia to Age- and Onset-matched comparison groups, since all participants groups should be equally disadvantaged by any collocations specific to British English to similar degrees, but we acknowledge that it may impact comparisons between our data and the original Gyllstad dataset.

Future Directions

The results reported here expand both the breadth of language representations supported by hippocampal-dependent memory as well as our understanding of the duration of hippocampal involvement in maintaining language representations after their initial acquisition. Our data provide cross-sectional evidence for the necessity of the hippocampus for supporting recognition of common English collocations. Future longitudinal designs that probe remote semantic representations over time after the onset of amnesia would strengthen the conclusions of this study. Another interesting question is how collocation knowledge changes over time in healthy adults. Such studies would require more sensitive measures of collocation knowledge, including greater variability in collocation frequency and entry into the English language compared to the tasks utilized here, which are comprised of high-frequency collocations drawn from the BNC, which is itself comprised predominantly from written material from the 1960s though 1990s (Gyllstad, 2007; Aston & Burnard, 1998; Meyer, 2002). Future studies could also examine production of collocations in the discourse of individuals with hippocampal amnesia. Data from such a study would enrich existing findings illustrating the impacts of hippocampal pathology and impoverished semantic representations on language production (Hilverman et al.. 2017; Rosenbaum et al., 2009; Verfaellie et al.,. 2014).

Here, we tested patients with bilateral hippocampal damage and frank¹ amnesia, but our results suggest that maintenance of remote semantic representations might be vulnerable in any population characterized by hippocampal abnormality, including conditions typically characterized by primarily by memory impairment (e.g. Alzheimer’s dementia, traumatic brain injury) and those typically characterized primarily by language impairment (e.g. specific language impairment/developmental language disorder). The prevalence of hippocampal dysfunction across a range of developmental and adult-onset disorders suggests that future research is needed to develop assessments sensitive to a range of language and communicative ability and able to detect hippocampal-dependent language impairments.

Conclusions

The current study provides new evidence that the hippocampus contributes to ongoing maintenance and strengthening of linguistic representations. In L1 speakers of English, hippocampal pathology impairs recognition of common English collocations, suggesting that the hippocampus is necessary for maintaining robust representations of this co-occurrence information. Our work answers calls to increase cross-talk between memory and language researchers interested in forgetting (in memory) and attrition (in language) (Mickan et al., 2019), and contributes important insights relevant to theories of language learning in adulthood, L1 language attrition, and memory consolidation.

Data Availability Statement

De-identified numerical item-level data from the experimental tasks and R scripts used in the statistical analyses are publicly available on OSF (see link in text). The conditions of our ethics approval do not permit public archiving of the clinical assessment and MRI study data. We note that all these data exist in various forms in multiple published papers, cited in the text. Readers seeking access to the data should contact author Melissa Duff. Access will be granted to named individuals in accordance with ethical procedures governing the reuse of clinical data, including completion of a formal data sharing agreement and approval of the local ethics committee.