The hippocampus supports the representation of abstract concepts: Implications for the study of recognition memory

Words, unlike images, are symbolic representations. The associative details inherent within a word ’ s meaning and the visual imagery it generates, are inextricably connected to the way words are processed and represented. It is well recognised that the hippocampus associatively binds components of a memory to form a lasting representation, and here we show that the hippocampus is especially sensitive to abstract word processing. Using fMRI during recognition, we found that the increased abstractness of words produced increased hippocampal activation regardless of memory outcome. Interestingly, word recollection produced hippocampal activation regardless of word content, while the parahippocampal cortex was sensitive to concreteness of word representations, regardless of memory outcome. We reason that the hippocampus has assumed a critical role in the representation of uncontextualized abstract word meaning, as its information-binding ability allows the retrieval of the semantic and visual associates that, when bound together, generate the abstract concept represented by word symbols. These insights have implications for research on word representation, memory, and hippocampal function, perhaps shedding light on how the human brain has adapted to encode and represent abstract concepts.


Introduction
The perceptual features of the objects around us, and scenes we are part of, can be encoded as visual representations using an approximation to a one-to-one mapping process.However, word-based information, which is a more recent addition to our information processing, cannot be meaningfully represented in this way.The perceptual features of a word (i.e., letters or sounds), unlike those of an object or scene, cannot generally form the basis of a representation that holds meaning or endures in memory.Instead, the conceptual meaning, or semantics, as well as the perceptual details that a word brings to mind, are what form the basis of an enduring representation accessible to memory.Exactly how the brain supports this remains unclear.Moreover, exactly how word information is represented in the brain can impact the type of memory generated and determine the specific brain regions, or networks, involved (Bartlett, 1932;Brandt et al., 2005;Kafkas and Montaldi, 2018a;Long and Prat, 2002).
For many reasons, often practical ones, word stimuli have been used very frequently to explore the neural basis of memory.In most cases, findings from word-based studies have been generalised to all forms of encoded information and have informed our current understanding of the neural basis of memory.Whether this generalisation is justified remains to be established.For example, findings from investigations into the way words, representing different types of concepts (e.g., abstract versus concrete) are represented in the brain (Hoffman et al., 2015;Vigliocco et al., 2014) may have critical implications for what the recognition data, acquired using word stimuli, can tell us about theories of memory.
The current study was motivated by the findings of an ambitious fMRI experiment (Mayes et al., 2019), where we aimed to compare the neural bases of recollection and familiarity, when the amounts of information processed, and the strength of memory reported, were matched.To do this we used a cued-recall word-completion recollection task,combined with our parametric familiarity paradigm (Kafkas et al., 2017;Montaldi et al., 2006) with word stimuli for the first time.While the cued-recall results confirmed that hippocampal activity was sensitive to amount recalled, but not to strength/confidence, most surprisingly, the familiarity results suggested that the hippocampus was sensitive to the strength of word familiarity.While consistent with earlier studies by Smith et al. (2011) and Song et al. (2011), this finding was strikingly inconsistent with all our previous studies using this familiarity paradigm (Kafkas et al., 2017;Kafkas and Montaldi, 2012;Montaldi et al., 2006), where non-word stimuli had been used.We hypothesised that this difference, rather than presenting us with an unexplained conundrum, might be highlighting the existence of a specialised role played by the hippocampus in the representation of words, and their meaning (Mayes et al., 2019).
We asked, therefore, how might the hippocampus play such a special role in supporting memory for word-based information, and what is it about words that makes them draw on the particular functions of the hippocampus.Answering these questions will provide initial evidence as to whether the hippocampus has been adopted to support the representation of words in a way that is not needed for other kinds of information.Using fMRI, we explored the neural mechanisms that support the representation and retrieval of words, while varying their abstractness.Abstractness describes the degree of visual imagery generated by a word and how tightly defined or not the meaning generated by that word may be (Barsalou and Wiemer-Hastings, 2005;Troche et al., 2014).We hypothesised that as memory for an abstract word cannot draw on a rich perceptual representation, and must require the bringing together, or association, of its key defining conceptual features, the neural mechanisms that support the representation of, and memory for, words might differ as a function of word abstractness.To better understand the mechanism that engages the hippocampus in word representation and memory, we exploited the algorithmic differences inherent in the familiarity/recollection distinction.We were therefore able to explore the effects of abstractness on both these forms of recognition memory, and on their neural bases, as well as the effects of abstractness independent of memory outcome.

Participants
In total, 21 participants gave informed consent and participated in the study.From these, 3 participants were excluded from the analyses; 1 due to a technical problem preventing the recording of fMRI data at retrieval and another 2 due to chance performance in the memory task.Therefore, the final sample included 18 participants (4 male), with a mean age of 22.78 years (SD = 2.21).The sample size was informed by our previous experiment utilizing word stimuli with a similar experimental procedure (Mayes et al., 2019).Also, a power calculation (using NeuroPower; URL: http://neuropowertools.org/) using data from our previous study (parametric familiarity contrast reported in Mayes et al., 2019), determined that for 18 participants we gain a power of 60%, which is considered moderate and therefore acceptable for the aims of the present study.All participants were native English speakers, right-handed, neurologically healthy and had normal (uncorrected) vision.The study received ethical approval from The National Research Ethics Service (North West-GM South) and each participant was paid £20 after completing the testing session.

Materials
Stimuli consisted of 402 word nouns (plus 14 words for practice) obtained from the MRC Psycholinguistic Database (Coltheart, 1981).Abstract words were selected to have a concreteness rating lower than 400 (range = 204-371; mean = 309, SD = 37), while concrete words were selected to have a concreteness rating higher than 500 (range = 516-653; mean = 575, SD = 29).Overall concreteness rating was significantly different between abstract and concrete words (t 400 = 79.71,p < 0.001).Concrete and abstract words were selected to have MRC familiarity ratings (indicating general level of encounter with the word) of at least 400 (abstract range: 420-612; concrete range: 421-657) and the two word groups were matched for familiarity rating (t 400 = 1.11, p = 0.27), Kucera-Francis written frequency scores (t 400 < 1) and word length (range = 5-9 letters).

Experimental procedures
Each experimental session consisted of two phases: an encoding phase completed outside the MRI scanner and a retrieval phase completed in the scanner (Fig. 1a).At encoding participants studied a series of 268 words, randomly selected from a pool of 134 concrete, and 134 abstract.Each word was presented centrally, in black letters (font: Arial 25 pt) on a white background for 3s.For each word participants were asked to make a frequency decision indicating how often they encounter each word in their everyday life (either spoken or written) using a binary decision ("quite often" or "not so often").The selection of the encoding task was informed by previous unpublished piloting, which established that this task results in a good response spread across the familiarity memory rating scale and overall adequate memory performance.Three other encoding tasks (involving pleasantness, alphabetic and semantic decisions) were rejected as they resulted in either poor overall memory performance (alphabetic decision) or in memory decisions restricted to the high end of the rating scale (either strong familiarity or recollection; pleasantness and semantic decision tasks).
After the encoding block and before starting the MRI scan, participants were trained in using a modified version of the remember/know procedure (Kafkas et al., 2017;Kafkas and Montaldi, 2012;Mayes et al., 2007;Migo et al., 2012;Montaldi et al., 2006).This entailed discriminating instances of familiarity, using a 3-point rating scale (from weak to strong familiarity: F1 = weak; F2 = moderate; F3 = strong familiarity), instances of recollection (R) and new stimuli (N).Participants were instructed to provide a familiarity response when they felt that they had encountered a stimulus at study, but report words as recollected if they spontaneously recalled additional associative information from the study episode in relation to a stimulus.After the training and to ensure understanding of these instructions, participants were asked to explain familiarity and recollection, to provide examples from their own experience and to ask questions about these two kinds of memory (for a detailed explanation of the procedure see Montaldi and Kafkas, 2024).A practice block that resembled the retrieval task was completed outside the scanner and one more practice was given in the scanner before the main task.
At retrieval, inside the scanner, participants were presented with 268 abstract and concrete words from encoding along with 134 new foils (67 concrete and 67 abstract).The fMRI data were collected and analysed from this phase and the session was divided into 2 functional runs.Each word was presented for 3s and participants were asked to use a 5-button MR-compatible button box to report whether each stimulus felt familiar, using three levels of increasing familiarity (i.e., F1, F2 or F3), whether it is recollected or new.Three fingers on one hand were used to select the familiarity responses and two fingers on the other hand were used for the extra two options (R and N).This allocation to the right or left hand were counterbalanced across participants.
Each word trial started with a fixation cross presented for 800ms and was followed by a mean intertrial interval of 1s (between 0.8 and 1.5s).
For each trial, the category of word (abstract or concrete) was chosen randomly, ensuring that no more than five trials of the same category (abstract/concrete) occurred consecutively.This sequence was organized in a counterbalanced manner for optimal efficiency, utilizing the Optseq2 software (https://surfer.nmr.mgh.harvard.edu/optseq/).A set of 81 implicit baseline (null) events (3.8s duration) were intermixed with the word stimuli (optimised using Optseq2), with no more than two null event presented consecutively, to ensure effective jittering and to provide baseline BOLD measures.During the fMRI session, participants were provided with earplugs and ear protection headphones and soft pads were used to stabilise their head to prevent movement artefacts.

fMRI data acquisition and pre-processing
A 3T Philips Achieva scanner was used to acquire the MRI data.
Functional data (gradient echo-planar images; EPI) were acquired using the blood oxygenation level dependent (BOLD) contrast and a total of 735 vol (TR = 2.5s; TE = 35ms; 40 slices per volume; matrix size = 96 × 96; voxel size = 2.5 × 2.5 × 3.5 mm) were recorded for each participant across three sessions covering the whole brain.T1 high resolution images were also collected from each participant at the beginning of each session, before the functional blocks (180 slices; voxels size = 1 mm isotropic; matrix size = 256 × 256).
The ArtRepair software (http://cibsr.stanford.edu/tools/human-brain-project/artrepair-software.html) was used to examine data quality of the EPI time-series and residual movement parameters were used in the GLM models (see below).SPM12 (Statistical Parametric Mapping, Wellcome Trust Centre for Neuroimaging; http://www.fil.ion.ucl.ac.uk/spm/) was used for data pre-processing and analysis.Data preprocessing included realignment of the EPI data to the mean image Fig. 1.Design of the fMRI study and behavioural results a) At encoding a series of abstract and concrete words were studied, while word frequency judgements were made.At test, while scanned, participants engaged in a recognition memory task in which previously studied words could be reported as recollected (i.e., retrieval of associated details related to the study episode), or familiar (i.e., words were recognised as being studied at encoding without triggering recall of additional details from study).Familiarity was reported using a rating scale (weak: F1, moderate: F2, strong: F3 familiarity), while separate and single responses were used to report recollected (R) and new/unstudied (N) stimuli.b) Proportions of hits (accurate old responses; bars) and false alarms (FA; incorrect old responses; lines) across the different memory outcomes plotted separately for abstract and concrete words.c) Memory performance for successful recognition of old stimuli (F1, F2, F3 and R) for abstract and concrete words, calculated as proportion of hits minus proportion of false alarms.Error bars indicate the standard error of the mean across participants.
using a standard six-parameter rigid body transformation, reslicing using sinc interpolation and slice timing correction (to the middle slice).Each subject's high resolution T1 image was coregistered to the corresponding mean EPI image.Finally, the coregisterd T1 and functional images were spatially normalised to MNI space using the DARTEL tool in SPM12 (Ashburner, 2007).DARTEL normalisation has been shown to provide improved realignment of the MTL structures across multiple participants when a whole-brain analysis is conducted (Yassa and Stark, 2009).After spatial normalisation the functional data were resliced to 3 mm isotropic and were spatially smoothed, for the univariate data analysis, using a 6 mm isotropic full width half maximum (FWHM) Gaussian kernel.

Univariate fMRI analyses
The pre-processed fMRI data were analysed using the general linear model (GLM) separately for each participant at the first level analysis.The onset (and the duration) of each event of interest was convolved with the canonical haemodynamic response function.For each participant, three models were defined, one trial-specific, one memory categorical and one memory parametric (see below).Regressors of no interest were also modelled and included trials with no behavioural response, the six movement parameters after realignment for each functional run and residual movement artefacts identified from the ArtRepair tool.A high-pass filter of 128s was applied to the data to remove low-frequency noise.
To explore BOLD responses as a function of stimulus abstractness irrespective of the type of memory response accompanying each word, a parametric model was constructed for each participant modelling all retrieval trials as separate events (trial-specific model).In this model, the abstractness rating of each word (obtained from the MRC Psycholinguistic database) was used as a covariate and was convolved with the trial-specific HRF.In a separate model, we specifically investigated parametric abstractness in recollected trials only, with all the other trials specified as a condition of no interest.First-level parametric t contrasts were created for linear increases as a function of increased concreteness or linear decreases as a function of increased concreteness mapping therefore brain responses to increased abstractness.Each of these firstlevel t-contrasts were used in the second-level (group) analysis implemented as one-sample t-tests.
To explore the memory interaction with abstractness, in the categorical model separate regressors for each condition of interest were defined for each participant.These consisted of all the potential memory outcomes separately for the two types of word (F1_abstract, F1_concrete, F2, abstract, F2_concrete, F3_abstract, F3_concrete, R_abstract, R_concrete, CR_abstract, CR_concrete, M_abstract, M_concrete, FA_collapsed).Trials with no responses were also modelled as conditions of no interest.In order to explore modulation of brain activity by familiarity strength for abstract and concrete words a parametric model was constructed for each participant (Büchel et al., 1998).In this model, familiarity hits (i.e., old stimuli reported as familiar) were specified as separate conditions, and the reported familiarity strength associated with each event was used as a covariate and was convolved with the onset-specific HRF.In this model, two parametric conditions were specified for the two word types (abstract and concrete words).In each parametric condition, four levels of strength were specified with misses (old stimuli deemed new) used as the level with zero familiarity (F 0 ), while F1, F2 and F3 responses were used as reflecting increasing levels of familiarity (from weak to strong).In this model, all other conditions (in addition to familiarity responses) were specified as regressors, following the same approach used in the categorical model.At the first (subject) level parametric t-contrasts were created for linear (monotonic) increases or decreases in activity across familiarity strength for abstract and concrete words (separately) and each of them were used in the second level (random effects) analysis implemented as one-sample t-test.Non-linear (quadratic) effects were also modelled to capture residual non-linear effects, but as these did not produce additional activations (not already included in the linear contrasts), they are not reported separately.
Regions of common modulation by familiarity strength for abstract and concrete words were examined using conjunction analysis based on the conjunction null hypothesis (Nichols et al., 2005).The parametric contrasts were also analysed using exclusive masking to explore activations unique to each type of stimulus (i.e., parametric familiarity for abstract by parametric familiarity for concrete words and vice versa; exclusive mask threshold: p < 0.05).This analysis reveals the brain regions that are modulated by familiarity strength for one type of word but not for the other (i.e., uniquely characterising familiarity for either abstract or concrete words).Recollection-related activations were also explored in the whole brain by contrasting recollections to abstract and concrete words with misses and F3 responses (i.e., recollections to abstract words: R abstract > M abstract and R abstract > F3 abstract ; recollections to concrete words: R concrete > M concrete and R concrete > F3 concrete ).Significance in all analyses was determined at a cluster-corrected family-wise error (FWE) p < 0.05 determined via nonparametric permutations as implemented within the Statistical NonParametric Mapping toolbox (SnPM 13; URL: http://warwick.ac.uk/snpm).

Regression analysis
Using the trial-specific GLM model, activation data (parameter estimates) for each trial and each participant were extracted from a hippocampal cluster identified in the whole-brain parametric analysis using a mask of 4 mm sphere around the MNI peak (− 30 -22 -14).Linear regression analyses were performed between the hippocampal activation data and degree of concreteness/abstractness (MRC Psycholinguistic rate of concreteness) of each word for the familiar trials (i.e., words which participants reported as familiar) as well as for all the words irrespective of memory outcome.

Multivoxel pattern analysis
Multivoxel patterns analysis (MVPA) was also conducted, on the spatially normalised but unsmoothed data, to further characterise the extent to which hippocampal activity successfully predicted (or classified) abstract and concrete words based on their reported familiarity or their property as abstract or concrete.Similar analyses were also conducted for the parahippocampal cortex and caudate nucleus, as these structures were also found to respond to stimulus reported familiarity (the latter are presented in Suppl.Analyses).Anatomical masks of the bilateral hippocampus, the bilateral parahippocampal cortex and the bilateral caudate nucleus were used as Regions of Interest (ROIs) derived using the PickAtlas Toolbox (Maldjian et al., 2003).The Pattern Recognition for Neuroimaging Toolbox (PRONTO, http://www.mlnl.cs.ucl.ac.uk/pronto/ (Schrouff et al., 2013) was used for the MVPA analysis.Classification accuracy in the ROIs was explored for familiar (collapsed F1, F2, F3 responses) abstract and concrete words relative to misses for each word type (i.e., familiar abstract vs. missed abstract and familiar concrete vs. missed concrete) using a set of binary support vector machine (SVM) classification procedures.Furthermore, classification success within the ROIs was also assessed for abstract vs. concrete words irrespective of the behavioural response associated with each response using a binary SVM algorithm (i.e., abstract words vs baseline and concrete words versus baseline).MVPA classification was also conducted for R events measuring accuracy in discriminating R vs. misses and R vs. F3, separately for abstract and concrete words, using again a set of SVM algorithms.All classification analyses were performed within the bilateral masks and separately for left and right lateralised masks and included all voxels within each ROI (no feature selection was used).The data were mean centred, and a leave-one-subject-out cross validation method was adopted to perform group analyses.Statistical significance of the classification outcomes A. Kafkas et al. (accuracies) for each model within the ROIs was tested using non-parametric permutations with 5000 iterations.

Behavioural results
Participants encoded a series of abstract and concrete words (134 words for each type) outside the scanner while making frequency decisions (see Fig. 1a and Methods).In the scanner, studied (268) and unstudied (134) words were presented and participants were trained to make recognition memory judgements for each word.At encoding, participants rated abstract words as 'quite often' 49% (SD = 20%) and 'not often' 50% (SD = 18%), while concrete words were rated 'quite often' 45% (SD = 15%) and 'not often' 55% (SD = 15%).

Contrasting abstract-sensitive and concrete-sensitive brain regions in the context of memory retrieval
First, we explored the brain regions that increase their activity as a function of concreteness or abstractness, independent of participants' memory judgment.A whole-brain analysis was performed on a trial-bytrial basis, with the degree of abstractness as the parametric modulator (see Methods; findings in Fig. 2 and Supplementary Table 2).Activity in the bilateral hippocampus, the bilateral insula (BA13), the left middle occipital and lingual gyrus (BA19) and the left middle cingulate (BA31) tracked increased word abstractness.On the other hand, decreases in abstractness (i.e., increases in concreteness) resulted in increased activity in the left parahippocampal cortex (BA36), the left precuneus (BA7/31), the left superior medial and lateral prefrontal cortex (BA6/8) and the left inferior parietal lobe (BA39/40).

The hippocampus is sensitive to the recollection of all words
The first analysis presented above established that the hippocampus responds to word abstractness irrespective of memory outcome.We next explored whether the hippocampus responds to recollection for both abstract and concrete words.Whole brain responses to recollection relative to misses and strong familiarity (F3) for both abstract and concrete words are shown in Fig. 3a and b (see also Supplementary Tables 3 and 4).A univariate analysis indicated that recollection relative both to misses and to F3 produced activations within the hippocampus for both word types (Fig. 3a and b).A separate parametric analysis, with abstractness as a parametric modulator on recollected trials only, did not produce any hippocampal effects (or effects in any other parts of the brain) either for increased abstractness or increased concreteness.Finally, MVPA classification analysis showed that activation within the hippocampus (bilaterally and separately for left and right) significantly discriminated recollected events from missed and F3 trials (Fig. 3c and  d); see also data in Supplementary Table 5).These findings indicate reliable hippocampal activations for recollection, which while diagnostic of recollection, are independent of abstractness.

The hippocampal response to word familiarity is driven by abstractness
We have established that the hippocampus responds to word abstractness irrespective of memory outcome, and that the hippocampal role in recollection is characterised by its activation for word recollection, independent of abstractness.Next, we turned to explore hippocampal sensitivity to abstract and concrete words when memory was supported by familiarity.First, parametric analyses explored how rated familiarity strength modulated regional brain activity, separately for abstract and concrete words (see Methods).Hippocampal and caudate nucleus activity were uniquely modulated by abstract word familiarity, while parahippocampal cortex (BA 35) activity was uniquely modulated by concrete word familiarity (Fig. 4a and b and Suppl.Table 6).Exclusive masking methodology confirmed that these parametric familiarity responses were selective for abstract words (caudate nucleus and hippocampus) and concrete words (parahippocampal cortex; Fig. 4a  and b and Suppl.Fig. 1).In addition to these abstract-unique and concrete-unique activations, these analyses also revealed a network of somewhat overlapping extra-MTL brain regions sensitive to abstract and concrete word familiarity strength (see Suppl.Table 6 and Suppl.Fig. 2).

Abstractness drives the hippocampal activity independent of familiarity
A regression analysis of the familiar words showed that the degree of hippocampal activation (from a 4 mm sphere centred around -30 -22 -14 as shown in Fig. 4a) significantly tracked the degree of word abstractness with lower activation for the most concrete words and higher activation for the most abstract ones (R 2 = 0.40, p < 0.001; Fig. 4d).This was also true when abstract (R 2 = 0.20, p = 0.05) and concrete words (R 2 = 0.25, p = 0.014) were assessed separately.Furthermore, BOLD activity in the hippocampus tracked the degree of abstractness across all words encountered at test, independent of what response was given and of whether previously encoded (Fig. 4e), with greater activation generated by the most abstract words (R 2 = 0.41, p < 0.0011).Again, this was also true when abstract and concrete words were examined separately (concrete: R 2 = 0.34, p = 0.004; Abstract: R 2 = 0.32, p = 0.001).Therefore, the hippocampus responds reliably to the degree of abstractness of words, and, in the context of memory retrieval, this is true not only in the case of words reported as familiar but for any word, irrespective of its memory status or the memory response it generates.
To identify any relationship between abstractness and familiarity ratings, the average abstractness rating of words rated as F1, F2 and F3 (as well as misses) was calculated for each participant (Fig. 4c).The mean abstract rating increased with increased familiarity memory strength (F 3,51 = 17.28, p < 0.001, η 2 = 0.50) with mean abstractness rating being significantly higher for F3 compared to each of the other levels (all ps < 0.01; see Fig. 4c).This strongly suggests that the familiarity-related activation in the hippocampus (Fig. 4c) was driven predominantly by abstractness.
Furthermore, with MVPA we examined whether cluster activity within the anatomical region of the hippocampus (bilateral and left/ right separately) could predict the accuracy with which familiar words were discriminated from missed words.As shown in Fig. 5a, activity within the bilateral hippocampus did not discriminate familiar from missed abstract words (accuracy = 38.9%,p = 0.84) or concrete words (accuracy = 50%, p = 0.57), and similar results were observed when examining left and right hippocampus separately (Fig. 5).However, when examining the hippocampal activity classification performance for all abstract and concrete words (irrespective of memory status or response), classification accuracy was above chance for abstract (accuracy = 77.8%,p = 0.014) but not concrete words (accuracy = 50%, p = 0.53; Fig. 5b).The effect for the abstract words was evident in the left hippocampus (accuracy = 72.22%,p = 0.03) but not the right hippocampus (accuracy = 55.6%, p = 0.41).Therefore, the MVPA analysis demonstrates that this hippocampal activity does not reflect familiarity memory but instead, is sensitive to the abstractness of the words.
Further support for the selectivity of this hippocampal sensitivity to word abstractness was provided by an MVPA analysis of the caudate nucleus (Fig. 5e and f) and the parahippocampal cortex (Fig. 5c and d), regions that showed selective univariate responses to familiar abstract (caudate) and familiar concrete (parahippocampal cortex) words (Fig. 4b and Suppl.Fig. 1).These analyses showed (see Fig. 5 and Supplementary Analyses) that unlike the hippocampus, activity within these regions was driven predominantly by familiarity memory and not by the degree of abstractness of the words.

Discussion
The aim of the present study was to evaluate the way that conceptual meaning and imagery which underpin the processing of verbal material, may affect the neural systems engaged during memory retrieval.Specifically, the study was motivated by the observation that the hippocampus appears sensitive to word familiarity (Mayes et al., 2019) but not familiarity for other types of information (Kafkas et al., 2017(Kafkas et al., , 2020;;Mayes et al., 2019).We asked what might be particular about words that they draw on the hippocampus to support a kind of memory (i.e., familiarity) which is described as being largely non-associative, and almost universally, as non-hippocampal (for discussion see Montaldi and Kafkas, 2024;Montaldi and Mayes, 2010).We reasoned that answering this question would be informative in understanding inconsistencies in the data regarding the role of the hippocampus in recognition memory.We further reasoned that levels of abstractness may have important implications for the way words are processed (Hoffman, 2016) when memory encoding and retrieval take place, and we therefore manipulated this variable to address our questions.
While, as expected, we found the hippocampus to be highly sensitive to word recollection (whether the word was high or low in abstractness), we strikingly also found that hippocampal activity appeared to systematically track the level of familiarity of words.We found, however, that this univariate activation in the hippocampus, as a function of the degree of reported familiarity, is confounded by the processing of word material.The hippocampal activity identified in the univariate analysis did track the degree of abstractness of familiar words, but critically, also the abstractness of all words used at test, irrespective of memory status or reported memory outcome.Furthermore, the MVPA methodology confirmed that the hippocampal BOLD response could not accurately discriminate familiar from unfamiliar (missed) words.Instead, the signal could accurately discriminate abstract from concrete words., d, f).For the familiar word analyses, binary classification success was calculated compared to misses (i.e., familiar abstract words vs. missed abstract words and familiar concrete words vs. missed concrete words).For the abstract and concrete analyses, binary classification success was calculated versus baseline activity (i.e., abstract words vs. baseline and concrete words vs. baseline).Significance was assessed based on permutation testing with 5000 permutations.*p < 0.05; **p < 0.01; ***p < 0.001; ‡p = 0.056/0.054(trend).
Error bars indicate the standard error of the mean across participants.
Finally, the hippocampal BOLD response was diagnostic of recollection over and above its response to abstractness.Interestingly, we also found a systematic tendency of participants to report abstract words as more familiar, which is consistent with the increased false alarm rates also found with abstract words.
These data, therefore, show that word abstractness, rather than word familiarity, modulates, and is predicted by, hippocampal activity.This suggests that, at least under some conditions, when words are encountered, the hippocampus is recruited to help form the associative representation of word meaning, when that meaning is somewhat, or very, abstract, irrespective of whether that representation underpins a memory or not.We believe that this particular hippocampal role aids the creation and evaluation of a representation by retrieving the semantic and conceptual associates of a word without triggering recollection of the word itself.Indeed, there is growing evidence suggesting involvement of the hippocampus specifically in language processing.For example, theta oscillations in the hippocampus increase while participants process sentences with constrained, relative to unconstrained, meaning (Covington et al., 2016).Also, successful linguistic prediction of word categories during sentence comprehension has been found to implicate the hippocampus along with other regions of the language network (Bonhage et al., 2015).Moreover, an interesting study has shown that patients with MTL damage that includes the hippocampus, while performing normally on standard neuropsychological tests of semantics and language, are impaired on vocabulary depth and richness (Klooster and Duff, 2015).This indicates that the hippocampus may play a role in some aspects of abstract concept processing, as recent studies suggest (e.g., Harpaintner et al., 2020).However, the reason the hippocampus does not always feature as a primary region contributing to abstract concept representation in the semantic memory literature (e.g., Bucur and Papagno, 2021) remains to be explored.We speculate that this relates to the multifaceted nature of abstract concepts and the experimental settings in which it has been explored (Borghi et al., 2018).It remains clear, however, that specifying the conditions under which abstractness engages the hippocampus requires further investigation.
The established role of the hippocampus in forming and retrieving associative memories (Eichenbaum et al., 2007;Kafkas and Montaldi, 2018b;Maguire et al., 2015;Mayes et al., 2007) is not as inconsistent with the current findings as might first be thought.This associative mechanism is likely to facilitate the processing of verbal information when processing demands require the retrieval of semantic associates, as with abstract words.Indeed, according to the semantic memory framework, proposed by Crutch and Warrington (2005), abstract words are organised in terms of associative properties and relationships, and more so than are concrete ones.They argue that the representational systems that support abstract and concrete word meanings, have qualitatively different properties.Consistent with this is the argument that abstract words are relatively unconstrained perceptually or spatially, and are linked with diverse contexts relative to concrete words (Barsalou and Wiemer-Hastings, 2005).Indeed, from a linguistic perspective, it is argued that abstract words show high levels of semantic variability and complexity and are associated with a wide variety of linguistic and semantic contexts (Borghi et al., 2017;Hoffman et al., 2013), including for example, those characterised by emotionality, social interaction, morality, time, and space (Troche et al., 2014).From a neural perspective, therefore, if the cognitive substrates of abstract concepts are somewhat determined by their semantic content, their neural bases might be expected to vary more than for concrete words (Wilson-Mendenhall et al., 2013).Indeed, brain activation prediction models, derived from semantic attribute combinations, have been found to successfully predict concrete words, but fail to predict abstract words (Fernandino et al., 2015).This finding is consistent with the limited agreement reported between participants when providing definitions of abstract, compared to concrete, words (Goetz et al., 2007).
Differences between our processing and experience of concrete and abstract words is also reflected in the behavioural performance of our participants, and explains why, what might at first appear to be a hippocampal sensitivity to familiarity, is instead driven by its sensitivity to abstractness.Critically, the increase in strong familiarity and false alarms, reported by participants in response to abstract compared to concrete words, indicates that participants were more likely to attribute feelings of memory to familiarity than recollection in the case of abstract words, even when memory was illusory.Furthermore, the strength of reported familiarity increased with increasing abstractness (Fig. 4c), while hippocampal activity did not discriminate familiar from unfamiliar (i.e., forgotten) words, but abstract from concrete ones (Fig. 5a  and b).Here we propose that the abstractness of a word triggers the spontaneous generation of hippocampally-dependent associations, that when bound together produce the word's conceptual representation, and the more associations that are generated, the richer the representation.This process draws on the same hippocampal mechanism that supports episodic associative memory (Mayes et al., 2007(Mayes et al., , 2019)), and it draws on it increasingly, with increasing levels of abstractness.We therefore argue that since the level of activity in the hippocampus reflects the amount of information it is associating (Mayes et al., 2019), the more abstract a word, the more its representation depends on the hippocampus.In turn, as both hippocampal activity and associative processing increase, so does the richness and feeling of familiarity of the representation experienced.Most importantly, the product of the hippocampal signal, while actually reflective of associative processes, is interpreted by participants as familiarity.This is because the hippocampal signal reflects the generation of the concept that the word itself represents, and not extra-stimulus associative detail, memory for which would be reported as recollection.
In contrast, the processing of concrete words relies more directly on visual information (Paivio et al., 1968) involving the retrieval of stored representations of the visual and conceptual features that concrete words represent (Hoffman et al., 2015;Paivio et al., 1968;Wang et al., 2010).As the parahippocampal cortex has a well-recognised role in processing complex visual information (e.g., Bonner et al., 2016;Sato and Nakamura, 2003), its activation for concrete words found in the current study is not surprising.Indeed, the parahippocampal activity, tracks the concreteness of words, independent of memory.Importantly, it also predicts familiarity, which likely reflects the contribution detailed visual information makes to the richness of a representation and the subsequent feeling of familiarity attributed to it.Therefore, we argue that these two MTL structures fulfil very different word-related functions; the hippocampus supports abstract word representation, as well as word recollection, while the parahippocampal cortex supports concrete word representation, as well as word familiarity, but not word recollection or abstract word representation.Therefore, the functions of these structures combine to produce rich word representations but diverge to support different components of word recognition memory.In the present study, we focused solely on word stimuli varying in abstractness.However, future endeavours may expand to include other visual materials (e.g., pictures) that vary in abstractness, to further explore the specificity of the effects related to word stimuli.
While we interpret the hippocampal sensitivity to word abstractness as independent of memory, the experiment itself involved a memory paradigm, and therefore participants were engaging in memory judgements, irrespective of whether words were identified as old or new, correctly, or incorrectly.The important implication here is that during scanning, all words were subjected to a memory evaluation which will have triggered mechanisms that either assessed the familiarity of each word or retrieved word representations from memory.Critically, however, even words that were not in memory, (i.e., not studied prior to the scan), showed hippocampal sensitivity to abstractness.So, while this effect reflects not only the formation of word representations, it also reflects the detailed evaluation of a word representation.This is important to highlight because it might be that under some conditions, the use of abstract words can be achieved with limited, or no hippocampal engagement, but perhaps when the generation of the representation in its richest form is required (perhaps to maximise memory accuracy, or as in Klooster and Duff (2015), to process words in depth), then the hippocampus might become critical.This may also explain why the hippocampus does not always feature as part of the core semantic network that supports the representation of abstract concepts.These findings have important implications for our understanding of memory and amnesia and potentially also our ability to detect early hippocampal deterioration (as seen in some dementias).Therefore, future research may explore differences in processing abstract (compared to concrete) words in patients with developmental amnesia (DA), adult-onset hippocampal lesions, mild cognitive impairment, and healthy ageing.
Related to DA, in which semantic learning is considered to remain relatively intact (Elward and Vargha-Khadem, 2018), we hypothesise that there might be some deficits in the representation of abstract concepts, especially as the level of abstractness increases.As Duff et al. (2020) discussed, even among DA patients, the rate of learning is slower, the amount of information acquired is less, and there is a reduced capacity for generalization compared to controls (Gardiner et al., 2008;Elward and Vargha-Khadem, 2018).The latter is an important feature of abstract thinking, and perhaps abstract concept learning and representation.Therefore, these more subtle deficits in the domain of semantic learning are not inconsistent with the hypotheses suggested by our findings.From an evolutionary perspective, our novel findings may also reflect an adaptation of a memory system originally optimised for more concrete representations, where an approximate one-to-one mapping between object and representation can be supported by non-hippocampal structures in the ventral visual stream.We suggest that as human language has evolved from that dominated by the more basic object representations (tools, objects, faces), to incorporate more concepts, generally associated with social organisation, morality, and emotion, our brains have adapted to support their representation.This proposed adaptation draws on the hippocampus and its highly specialised associative mechanisms.Without these mechanisms we predict that the brain is only able to robustly represent abstract concepts and words in their simplest, or most degraded form.
At first glance, our findings may seem to conflict with those reported by Clark et al. (2018), where greater hippocampal activation was observed for object and scene word pairs compared to abstract word pairs during encoding.However, a critical distinction lies in the nature of the tasks used, with our study employing item recognition while Clark et al. (2018) used a paired associate task.This fundamental difference necessitates the use of distinct memory strategies, which in turn may trigger distinct conceptual and perceptual representations of the words involved.Notably, Clark et al. contrasted abstract word pairs with limited associative connections (e.g., Principles -Attempt), with the associative properties of object (e.g., Broccoli -Headphones) and scene word pairs (e.g., Orchestra -Boardroom).Clark et al. scanned during encoding when participants were instructed to actively associate the paired stimuli.It seems plausible that presentation of the object and the scene word pairs spontaneously triggered binding rich perceptual associations, used to encode the items as pairs, while the abstract word pairs will have triggered binding, but much less informationally rich than the other conditions.This argument was not tested by Clark et al. but is supported by their finding that abstract word pair memory was significantly worse than object or scene word memory at test.We know that the hippocampus is more sensitive to the amount of, rather than the confidence in, information recalled (Mayes et al., 2019).Therefore, a question arising from Clark et al., that requires further exploration, is whether matching amount of information recalled across word pair types would have removed their hippocampal difference.Taken together, while these two studies appear similar, and their comparison is interesting, their differences are fundamental.
Our findings provide novel observations of a hippocampal contribution to word processing as the degree of word abstractness increases.We propose that this reflects greater associative processing, equivalent to that which underpins recollection, but critically here, without the experience of recollection.Importantly, the hippocampal activity we report, is not limited to very abstract words, but is evident across the abstract-concrete spectrum and potential effects are, therefore, not limited to abstract words.These findings and their interpretation potentially explain inconsistencies, especially in previous fMRI studies, regarding the role of the hippocampus in recognition memory.We suggest that the presumed role of the hippocampus in familiarity memory (e.g., Smith et al., 2011) is illusory and is driven by the symbolic nature of words, requiring the generation of semantic associates when they are processed.The findings warn that the processing of verbal material, in general, may be under certain conditions more prone to engage the hippocampus than are other forms of material, even though recollection may not be consciously experienced.Finally, we speculate that our representational system may have adapted to meet the needs of a more sophisticated, complex, and abstract language.

Declaration of competing interest
None declared.

Fig. 2 .
Fig. 2. Whole-brain responses to degree of (a) abstractness (red) and (b) concreteness (blue) of the words.The illustrated regions are characterised by increased BOLD as a function of increased word abstractness (in red, upper panel) or increased word concreteness (in blue, lower panel) irrespective of kind of memory reported.HC = hippocampus; PHC = parahippocampal cortex; PrCu = precuneus; pCing = posterior cingulate; mSFG = medial superior frontal gyrus; IPL = Inferior parietal lobe; SFG = superior frontal gyrus; L = left; R = rightActivations are displayed at a voxel-wise p < 0.001 and are significant at a cluster-corrected FWE p < 0.05 determined via nonparametric permutations (all ts > 3.75).

Fig. 3 .
Fig. 3. Hippocampal responses to recollection of abstract and concrete words: Activations are strongly diagnostic of memory.a-b) Hippocampal activations for abstract (red; a) and concrete (blue; b) words reported as recollected relative to misses and F3 responses.R > M for abstract: MNI: 30 -13 -20; R > M for concrete: MNI: 15 -25 -17, cluster includes hippocampus and parahippocampal cortex.R > F3 for abstract MNI: 21 -7 -14, (cluster centred within the amygdala including 9 voxels in the anterior hippocampus); R > F3 for concrete MNI: 24 -28 -5).Activations are displayed at a voxel-wise p < 0.001 level and are significant with a clustercorrected family-wise error (FWE) p < 0.05 determined via nonparametric permutations.c-d) Classification (MVPA) outcomes within the hippocampus showing that hippocampal activity discriminates recollections (relative to misses and F3) for both abstract and concrete words.In separate analyses, binary classification success was calculated for recollected versus missed (c) and recollected versus F3 trials (d).Significance of classification success from activation data was assessed based on permutation testing with 5000 permutations.*p < 0.05; ‡p = 0.056/0.054(trend).Data (percent accuracy and p-values) are also presented in Supplementary Table 5. Error bars indicate the standard error of the mean across participants.

Fig. 4 .
Fig. 4. Activation increases in the MTL as a function of familiarity memory.a) Hippocampal activity (MNI: 30 -22 -14) is modulated by abstract word familiarity strength (surviving exclusive masking by concrete word familiarity).b) Familiarity response in the left parahippocampal cortex (MNI: 22 -27 -20) selective for concrete words (surviving exclusive masking by abstract word familiarity).Activations are displayed at a voxel-wise p < 0.001 and are significant at a clustercorrected family-wise error (FWE) p < 0.05 determined via nonparametric permutations.The parameter estimates plotted in red (a) or blue (b) indicate which familiarity parametric response was selectively significant in each regioneither familiarity response to abstract or concrete words; the non-significant parametric response is plotted in gray.c) Mean word abstractness rating for each recognition response category (smaller numbers on y-axis indicate greater abstractness) demonstrate that increased abstractness is associated with increased familiarity strength.d) and e) Regression analysis between hippocampal activity in the cluster presented in (a) and the degree of abstractness (from highly concrete to highly abstract words) for words judged as familiar (d) and all words (e) irrespective of memory outcome.Note: M = misses, F1 = weak, F2 = moderate, F3 = strong familiarity hits, R = recollection.***p < 0.001; **p < 0.01; *p < 0.05.All error bars show the standard error of the mean.

Fig. 5 .
Fig. 5. Classification (MVPA) outcomes within the hippocampus (a, b), the parahippocampal cortex (PHC; c, d) and the caudate nucleus (e, f) for abstract and concrete words judged as familiar (a, c, e) and all word stimuli irrespective of memory status or response (b, d, f).For the familiar word analyses, binary classification success was calculated compared to misses (i.e., familiar abstract words vs. missed abstract words and familiar concrete words vs. missed concrete words).For the abstract and concrete analyses, binary classification success was calculated versus baseline activity (i.e., abstract words vs. baseline and concrete words vs. baseline).Significance was assessed based on permutation testing with 5000 permutations.*p < 0.05; **p < 0.01; ***p < 0.001; ‡p = 0.056/0.054(trend).Error bars indicate the standard error of the mean across participants.