In the current study, we examine the data originally collected by Klooster and Duff (2015) to test whether participants' impaired performance on the feature generation task can be understood in terms of a relational semantic deficit. To do this, we use a computational model of semantic representational structure to characterize the memory search processes engaged by the task. This allows us to examine the semantic relations between generated features and the target word, and the relations of the set of generated features to one another. We find substantial differences in the nature of semantic search between the two groups, which we interpret in terms of current theories of semantic and episodic memory search. While Klooster and Duff (2015) also characterized semantic deficits in two other tasks, feature generation performance was the most amenable to semantic analysis: its generative nature allowed us to examine the dynamics of search, and participants overall produced about five times as many responses in this task relative to the senses task.

A number of algorithms have been developed to construct semantic representational codes from either large text corpuses (Lund and Burgess, 1996; Landauer and Dumais, 1997; Jones and Mewhort, 2007) or from behavioral responses in a free association task (Steyvers et al., 2004). These are often referred to as vector space models of semantics, as each representational code in the system is a vector of numbers. While the numerical features that comprise the representations in these vector space models are rarely directly interpretable, they do provide a reference point for each word, such that words with similar features are situated near to one another in the vector space. This computational approach allows us to consider a sequence of responses in the feature generation task as a trajectory through an abstract semantic representational space. This trajectory can be characterized in terms of the semantic distance between the target word and the individual features generated by participants, and the distance of the generated features to one another.

Hills et al. (2012) used a similar approach to characterize performance on a semantic fluency task, in which participants are asked to provide examples from the semantic category “animal.” In their model, the vector representation of the most recent response was used as a retrieval cue to determine the next response. The likelihood of recalling a particular word in a search of the category semantic space was proportional to its representational similarity to the most recent response. This framework naturally explains the semantic clustering seen in semantic fluency tasks: The initial response tends to be a highly frequent exemplar of the category (Henley, 1969; Newcombe, 1969), and the continual updating of the retrieval cue causes contiguous responses to be semantically similar to one another (Bousfield and Sedgewick, 1944; Federmeier et al., 2002; Voorspoels et al., 2014).

Our proposed model is similar to the Hills et al. model in that feature responses are based on a blended representation of target word and previous recall information. Critically, all of the HP patients in the (Klooster and Duff, 2015) study were impaired in feature generation. However, as a group they did not show a deficit in a measure of semantic fluency (the Controlled Oral Word Association test), although there seems to be more variability in their performance at the individual level. It is therefore important to understand how semantic feature generation differs from semantic fluency, and how the task demands might reveal the nature of semantic deficits in hippocampal amnesia. In both cases, the participant's knowledge is probed in a constrained way. With semantic fluency, responses are constrained to come from a particular taxonomic category (Gruenewald and Lockhead, 1980). In semantic feature generation, responses are constrained to be in reference to a target word, and are meant to describe properties or characteristics of the referent item. However, these tasks seem to require access to different kinds of conceptual representations. Firstly, the feature generation task cues semantic search with a more specific target than semantic fluency (e.g., “dolphin” vs. “animal,” respectively). Secondly, the task demands of feature generation requires retrieval of richer multi-word conceptual representations, whereas semantic fluency requires participants to name exemplars. Lastly, in semantic fluency each response is related to the others by the shared features that comprise category membership. In comparison, in semantic feature generation the adjacently retrieved features are potentially semantically unrelated outside the context of the target word (e.g., “gray in color” and “intelligent animal” for “dolphin”). This type of feature generation seems to require relational memory to access semantically disparate concepts that are related only given the context of the target word. Our semantic analyses suggest that this task distinction is important and can potentially unearth semantic memory deficits that are otherwise masked in surface-level tasks. In the discussion we consider the critical role of the hippocampus in relational memory, and the key differences between the tasks mentioned above and semantic tasks which are not associated with an impairment in patients with hippocampal amnesia. We will return to the question of how these results inform our understanding of hippocampal engagement in semantic memory search.

Participants were five patients with bilateral hippocampal damage (HP) exhibiting declarative memory impairments. Fifteen healthy participants (CO) were matched to the patient group on sex, age, and education (three matched participants to each patient). Each patient with hippocampal damage had stable, non-progressive lesions. The etiology of three patients was anoxia/hypoxia—resulting in bilateral hippocampal damage. Two patients had herpes simplex encephalitis, resulting in broader bilateral medial temporal lobe damage, including hippocampus, amygdala, and surrounding cortices. For more details see Klooster and Duff (2015).

In the feature generation task, participants were presented with a target word (e.g., “bed”) and given 2 min to verbally list as many features of that word as possible. Thirty-five target words were sampled from established feature production norms (McRae et al., 2005). Instructions and examples were given to the participant at the start of the task, and were left in front of the participants and repeated by the experimenter regularly. On each trial, the experimenter read the target word aloud, and prompted the participant to begin to report features. If, during this recall period, the participant stopped responding, the target word was repeated by the experimenter, and the participant was encouraged to keep trying to generate features. Responses were video recorded for later transcription and analysis.

Each participant's verbal responses were transcribed and two judges coded these responses into a sequence of features. See Klooster and Duff (2015) for details regarding this coding procedure. Table 1 provides representative examples of features. For the current study, we developed a coding scheme to include all content words and exclude function words from the semantic analysis. Excluded grammatical groups were: personal pronouns, possessive pronouns, auxiliary verbs (be, do, and have), coordinating conjunctions and articles.

In this study we used semantic representations constructed with the Global Vectors (GloVe) algorithm (Pennington et al., 2014), which has excellent coverage of the English language due to the large text corpus used to construct the vector representations. GloVe follows in a long tradition of computational models attempting to quantify the meaning of words by assigning each word a point in a high-dimensional vector space, often containing up to 300 dimensions (Deerwester et al., 1990; Lund and Burgess, 1996; Landauer and Dumais, 1997; Steyvers et al., 2004). These techniques tend to use linear algebraic algorithms (such as singular value decomposition) to construct vector representations given statistics characterizing the co-occurrence of words in a large text corpus.

These semantic vector space models formalize longstanding ideas from linguistics and philosophy regarding how best to characterize the meaning of words. In the linguistics literature, the Distributional Hypothesis refers to the notion that words that co-occur across similar contexts tend to have similar or related meanings (Harris, 1954). Linguist J. R. Firth famously summarized the context-dependent nature of meaning with the phrase “You shall know a word by the company it keeps” (Firth, 1957, p. 11). The assignment of vector representations to words and phrases resonates with ideas developed by Wittgenstein (1953), whereby words can be loosely grouped by a combination of shared features. Given these vector representations, the semantic similarity of two words can be quantified using standard distance measures such as Euclidean distance or the cosine angle between two vectors (Kwantes, 2005). In the current work we used the GloVe model to construct a single representation for multi-word features by taking the average of the semantic similarity of each of the feature words to the target word. For feature-to-feature analysis we calculated the pairwise similarity between each word in the two features; the average of these similarity scores was used to represent the similarity of the features to one another.

We implemented all Bayesian analyses in R. Initial analysis using a frequentist framework indicated that residuals were not normally distributed, motivating the use of a Bayesian framework. Further, the nature of cosine values is such that they are log-normally distributed, and a Bayesian framework gives us more flexibility to estimate this. To examine group differences in cosine similarity scores we fit a Bayesian linear mixed effects model using the Stan and brms packages in R (Bürkner, 2017; Stan Development Team, 2017). The binary group predictor (CO vs. HP) was modeled as a fixed intercept and slope. Subject (s) and target word (w) predictors were modeled as random effects with varying intercepts and normally distributed priors: s ~ Normal(0, σs), w ~ Normal(0, σw). Prior distributions on the variance parameters were uniform: σe, σs, σw ~ Uniform(0, ∞). We estimated the response distribution of cosine similarity scores as log-normal, as preliminary examination of the data showed that this distribution was better described with a log-normal distribution as compared to a normal distribution. We estimated model parameters using Markov-Chain Monte-Carlo (MCMC) methods, using the No U-turn Sampler (NUTS) provided with Stan. For all Stan-based model fits, we ran 4 chains each with 4,000 iterations to ensure chains effectively converged. Chain convergence was confirmed by the r^ statistic which in all cases approached 1 (indicating maximal convergence).

A second Bayesian linear mixed effects model was designed to characterize the cosine similarity of feature responses to one another (within a given response sequence). Similarity was calculated between features with lag one to four. Lag is defined as the positional difference in the response sequence, with adjacent features assigned a lag of one, features separated by one intervening feature assigned a lag of two, and so forth. Group, subject, and word predictors were modeled as described above, except that the response distribution of cosine similarity scores was modeled as normally distributed. Prior distributions on the variance parameters, MCMC details, and model comparison details were the same as above.

In order to examine changes in cosine similarity across the response sequence, and changes as a function of lag within a given response sequence, we created a set of Bayesian multilevel models. The data was best modeled by power functions, which take the form f(x) = a(x^b), where a is a scaling factor, and x is a variable base raised to a constant power, b. The b coefficient represents the growth or decay in cosine similarity scores as a function of x, which represents either response position or positional distance between generated features in different analyses. In our two-level hierarchical models, we estimate the group and subject-level effects of feature responses. At the group level, we estimated the a parameter for both groups with the prior μ ~ Normal(0.2, 0.5), σ ~ Cauchy(0, 5), and the b parameter with the prior μ ~ Normal(0, 0.5), σ ~ Cauchy(0, 5). We estimated model parameters using MCMC in Stan as described above. All chains converged effectively. As above, these models were compared to a null model without a group-level predictor.

Overall, the feature responses made by patients with hippocampal damage tended to be closer in semantic space to the target word (HP: μ = 0.21, SD = 0.13) when compared to healthy comparison participants (CO: μ = 0.19, SD = 0.13). This positive shift in the cosine distribution of HP responses can be seen in Figure 1. We used a Bayesian mixed effects regression framework to investigate the effect of group (HP vs. CO) on cosine similarity of feature response to target word. The model had a fixed effect of group and accounted for variance associated with individual subjects and target word stimuli. The posterior distribution of the MCMC chains for the group coefficient did not include zero (μ = 0.02, SD = 0.003, 95% CI = [0.002, 0.033]), consistent with a substantial and reliable group difference in cosine similarity. We considered the possibility that the increased cosine similarity of features to target word was driven by an individual patient, therefore we carried out a “leave-one-out” by patient analysis. We iteratively ran the model described above five times, each time excluding one patient's data. For each iteration, the resultant posterior distribution of the group parameter did not include zero, suggesting that no single patient was skewing the group result.

Firstly, we were interested in how participants initiate the search for features of a given concept in semantic space. We examined the cosine similarity between the target word and the first five feature responses. Participants in the HP group generally make fewer responses than the CO participants, but routinely make more than 5 responses. As such, restriction to the first 5 responses puts the two groups on relatively even footing in terms of the number of responses in each response position bin. In both HP and CO groups we found that the cosine similarity to the target word decreased across the first five responses (see Figure 2). While the first feature response of both groups was a similar distance from the target word, a group difference emerges over the course of the first five responses, with the CO responses ranging farther in semantic space on average relative to the HP responses.

To characterize these observed trends, we constructed a Bayesian hierarchical model, fitting power function curves to this sequence of response positions. Best-fit curves are presented in Figure 2B. The best-fit curves for the two groups have similar starting points (as reflected by the best-fitting a parameters, HP: 0.2536, CO: 0.2510), but the patient group shows a slower rate of decay in semantic similarity to the target word as the search progresses (as reflected by the best-fitting b parameters, HP: –0.1243, CO: –0.2018).

In order to characterize the statistical reliability of this difference in decay rates, we examined the MCMC-derived posterior distributions for each of the power function curve parameters. Intuitively, these posterior distributions contain the set of plausible parameter values for each of the groups. As we are interested in determining whether the shift in parameter values is reliable, we constructed a difference distribution for each of the parameters: Each sample in the posterior distribution specifies four numbers, the mean a parameter for the HP and CO groups (ā_HP, ā_CO), and the mean b parameter for the HP and CO groups (b̄HP,b̄CO). The difference distributions were constructed by calculating ā_HP−ā_CO and b̄HP-b̄CO for each sample in the posterior distribution.

For the a parameter, the mean of the difference distribution was near-zero (0.0027), with points tending to be evenly distributed around zero (in 57% of posterior samples ā_HP>ā_CO). This suggests that both groups initiate semantic search in a similar way. For the b parameter, the mean of the difference distribution was more substantially positive (0.0762, consistent with a shallower decay for HP), with 83% of the difference distribution falling above zero. In other words, the semantic relatedness of the generated features to the target word decayed more slowly for the HP group, consistent with the idea that the CO group is able to range further from the target word in semantic space. The group difference in the b parameter is consistent with the group difference established in the first analysis, but is less reliable statistically. This is likely due in part to the restriction of this analysis to the first five response positions, and also to the presence of fairly substantial individual differences, as can be seen in Figure 2B.

Considering the first two analyses together, it seems reasonable to infer that the group-level difference characterized in the first analysis is not present in the initial responses. This is consistent with a model in which the difference in semantic relatedness emerges over the course of the response sequence. A follow-up analysis showed that the mean target-to-feature semantic relatedness for the later response positions excluded from this analysis (through to the termination of the response sequences) is similar to the asymptotic values approached by the two power curves estimated here. These results are generally consistent with a model in which the hippocampus facilitates the retrieval of semantically distant features of the target word. We return to this point in the discussion.

The previous analyses examined the semantic similarity of the words comprising each feature to the target word specific to that trial. In order to better characterize the dynamic nature of semantic memory search, we calculated the semantic relatedness of the reported features to one another, without regard to the semantic identity of the target word. This allowed us to examine how feature-to-feature similarity changed as a function of the relative position of the two features in the response sequence. Figure 3 shows that as the positional lag between two responses on a given trial increases, there is a substantial decline in cosine similarity. In other words, as two responses become farther apart in the response sequence, they become less semantically related to one another.

We first used a Bayesian linear mixed effects model to estimate average feature-to-feature semantic relatedness, without considering transition lag. As before, the model had a fixed effect of group and accounted for variance associated with individual subjects and target words. The posterior distribution of the group coefficient was centered around zero (μ = 0.0068, SD = 0.0119), consistent with the idea that there is no HP vs. CO group difference in feature-to-feature similarity.

Once more, we constructed a Bayesian hierarchical model, fitting power functions to these curves to determine whether a group difference exists in the non-linear shift of semantic relatedness as feature responses become separated in the response sequence. Best-fit curves are presented in Figure 3B, with parameter estimates for each group in the caption. The best-fit curves for the two groups have similar starting points (the a parameter, HP: 0.2131, CO: 0.2077), and similar decay in semantic similarity as the positional distance between responses increases (the b parameter, HP: –0.1256, CO: –0.1184). As above, we calculated difference distributions, ā_HP−ā_CO and b̄HP-b̄CO, for each sample in the posterior distribution. For the a parameter, the mean of the difference distribution was near-zero (0.0059), with 65% of the samples in the CO group less than the HP group. For the b parameter, the mean of the b̄HP-b̄CO distribution was also near zero (0.0079) with 58% of the HP samples falling below the CO samples. This analysis suggests that the process governing transitions between generated features behaves similarly for the two groups.

We consider how semantic vector space representations could work with mechanisms commonly used to model search processes in episodic and semantic memory. In order to develop a model of the semantic feature generation task, we begin by comparing it to other memory search tasks that have been modeled computationally. In many theories, memory search is guided by the construction and utilization of a retrieval cue: a mental representation that targets and reactivates task-relevant memories. For example, in the semantic fluency model developed by Hills et al. (2012), a retrieval cue containing the most recently reported response was used to target local conceptual representations from the category “animal.” The representational similarity between the retrieval cue and each of the not-yet-recalled animals was used to simulate a decision competition in which the likelihood of a given animal winning the competition was proportional to its semantic similarity to the retrieval cue. The continual updating of the retrieval cue causes contiguous responses to be semantically similar to one another. Smith et al. (2013) used a semantic vector space model to examine the search process in a Remote Associates Test in which the participant must produce a target word that is semantically related to three presented cue words. Participants were encouraged to vocalize guesses as they attempted to determine the target word, and their model suggested a similar semantic dependence of a given response on the previous responses in the sequence. It is worth noting that hippocampal patients in the Klooster and Duff (2015) study were impaired on a similar Word Associates Test, in which remote semantic associates to a target word had to be identified.

In retrieved-context models of free recall, a retrieval cue comprised of context information is used to target episodic representations of words from a recently studied list (Howard and Kahana, 2002a; Sederberg et al., 2008). In many experiments, the temporal structure of the studied items dominates clustering during the recall period: items studied in nearby list positions tend to be recalled in adjacent output positions. In these models, recalling an item reinstates the context associated with that item at encoding which increases support for its neighbors at retrieval (Kahana, 1996; Kahana et al., 2008; Healey et al., 2019). There is a simultaneous influence of semantic relatedness on the order of recall responses (Howard and Kahana, 2002b; Polyn et al., 2009). As in the semantic fluency task, semantically related study items tend to be produced as contiguous responses in the recall sequence (Romney et al., 1993; Polyn et al., 2011). In a number of free recall models, these semantic organization effects arise from the dynamics of an ever-changing retrieval cue which integrates the representation of the just recalled item (Sirotin et al., 2005; Polyn et al., 2009; Socher et al., 2009; Morton and Polyn, 2016). Here we consider how similar mechanisms could be used to develop a model of the semantic feature generation task.

The current results provide constraints that can be taken into account in future modeling work. With regard to the functioning of the healthy cognitive system, we envision an executive system guiding task performance through the construction of a retrieval cue that probes a semantic memory space. This semantic memory space is populated with representations of known objects as well as representations of their features and characteristics. The retrieval cue activates a particular location in this semantic space, which activates nearby conceptual representations in proportion to their proximity to the activated location. This proximity-based activation is similar to the dynamics of a spreading activation model (Collins and Loftus, 1975; Anderson, 1983). These representations then compete to be retrieved, with their relative activity determining the support for each representation. The cosine similarity scores used in our analyses reflect the proximity of these representations to one another, and as such, can be thought of as approximating the level of support for each representation in this retrieval competition. The winning representation is fully activated, allowing that feature to be verbally reported. The retrieved feature representation can then be used to modify the retrieval cue, and semantic search continues.

The observed behavioral phenomena are consistent with this model. We propose that semantic space is probed and guided by a compound retrieval cue, containing a representation of the target word as well as a representation of the most recent feature response. The first features retrieved tend to be close in semantic space to the target word, suggesting that the initial search is guided by a retrieval cue that simply contains a representation of the target word. Subsequent responses range further from the target word in the semantic space, and neighboring responses tend to be more similar to one another than to other responses. One way for the system to support retrieval of more distant features in the semantic space is to integrate information related to already retrieved features into the retrieval cue itself, creating a compound cue of target and recent feature information. This compound cue would allow the system to target more distant parts of the semantic space, as features proximal to the already retrieved features would now receive additional support in the retrieval competition. By retaining target word information in the retrieval cue, the system can ensure that retrieved features remain relevant to the current target word. However, as the number of feature responses increases, the target word representation may become progressively less influential in the retrieval cue, allowing the system to range further from its point of initiation (as shown in Figure 4).

These results also provide constraints regarding the specific contribution of the hippocampus to semantic memory search, although there remain a number of open questions that we discuss in the following sections. Specifically, with hippocampal damage, feature responses have a restricted range in semantic space. However, at the same time the semantic relatedness of successively reported features to one another is unaffected. This raises the possibility that the executive machinery guiding search is unaffected, as it is still able to incorporate information about the previous response into the retrieval cue guiding search. Furthermore, patients are able to reliably stay “on task,” in that they consistently generate valid features of the target word. Indeed, Klooster and Duff (2015) found no significant group difference in the number of unrelated responses (p > 0.27) or the number of factually incorrect responses (p > 0.62) produced between the CO and HP groups. The deficit seems more specific to the patients' ability to access distant semantic features of the target word.

As reviewed in the introduction, the hippocampus has been clearly implicated in both relational processing and episodic memory. However, its role in semantic memory is less well characterized. We propose that people with hippocampal damage have difficulty using semantic knowledge in a flexible, relational manner. As mentioned above, neuropsychological studies have found that patients show minimal impairment in basic tasks probing semantic knowledge, but it is possible that these tasks mask a more subtle deficit in relational processing.

A number of studies indicate that the hippocampus contributes to successful relational memory – that is, the formation of long-term memories comprised of multiple elements bound together (Cohen et al., 1999; Konkel and Cohen, 2009). As we discuss below, relational processes can be independent of long-term memory and can refer to any cognitive mechanism involving relational representations. In the feature generation task the HP group shows an impairment in retrieving rich semantic representations: fewer features are produced, and the produced features do not range as far in semantic space as those produced by the CO group. We discuss possible explanations for this observed deficit. First, we consider that the relational search process, facilitated by a compound retrieval cue, is impaired. Second, we consider whether the deficit could arise directly from an impaired ability to retrieve episodic memories. Third, we explore the possibility that the HP group impairment is due to a general degradation of the semantic space used to represent features.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.