Our everyday experiences shape the way we conceptualize and act in the world. Following this intuition, previous work using text corpora has proved useful in understanding the nature of language and human concepts (Andrews et al. 2009; Deerwester et al. 1990). One appeal of this work is that text, such as from newspaper articles, reflects human behaviour outside the laboratory. However, this text primarily serves a communicative role and is often scraped from curated sources, making it less reflective of real human activity.
In this contribution, we aim to build upon previous work from the text domain by analyzing real-world behaviour from a broad section of the general population as they go about an everyday activity in relative anonymity, namely supermarket shopping. We apply techniques developed in computational linguistics to shopping data from nearly 1.3 million trips. Instead of words and documents, our analyses are over products and shopping baskets. These analyses reveal that human concepts are organized around goals and interactions (e.g. tomatoes go well with vegetables in a salad), rather than their internal features (e.g. defining a tomato by the fact that it has seeds and is fleshy).
Our work speaks to the relative importance of intrinsic and extrinsic features in concept representation. One way that people may reason about categories is to decompose them into intrinsic features or parts (Plato 1973). On this view, a bird is an animal that typically has wings, feathers, a beak, and so on (Rosch and Mervis 1975). However, extrinsic features are also critical for how humans organize concepts and come to understand the world, to the extent that some concepts may be solely defined by them (Barr and Caplan 1987). For example, Wittgenstein (1967) asserted that the concept of game is undefinable. One might suggest that games are fun, but Russian Roulette is not fun and other activities that are fun are not games. Likewise, not all games are competitive (e.g. Ring Around the Rosie). Instead of defining game in terms of intrinsic features, one solution is to define game relationally—a game is simply something that is played (Markman and Stilwell 2001). Human categories are therefore additionally sensitive to relationships and interactions with other concepts (Markman and Stilwell 2001).
The importance of relations and interactions extends beyond abstract concepts. Many features of concrete concepts are extrinsic (Jones and Love 2007). For example, whilst knowing that tomatoes are taxonomically related to fruits, people commonly associate them with other vegetables. Even for natural kinds, people commonly list extrinsic features for concepts (Jones and Love 2007), such as noting that birds eat worms. Meanings appear to update in light of extrinsic relationships. For example, people are more likely to judge a polar bear and a dog as similar after reading vignettes in which both played the same role in a relation, such as chasing some other animal (Jones and Love 2007). Likewise, merely sharing a thematic relationship, such as a man and a tie (e.g. wears), makes the linked concepts more similar (Schank and Abelson 1977; Wisniewski and Bassok 1999; Jones and Love 2007).
When concepts are defined in terms of other concepts, what moors or grounds our concepts to the physical world we inhabit (Harnad 1990)? One proposed solution is that some concepts are embodied (Barsalou 2008). For example, the action of hammering may be grounded to related motor programmes and associated perceptions, linking the body, mind, and physical world. Indeed, neuroscientific evidence has shown that comprehension of language is tightly coupled with the neural regions associated with action and perception (Pickering and Garrod 2013). A computational model developed by Mitchell et al. (2008) was able to accurately predict the neural activity elicited by a noun by considering the co-occurrence of that noun with action verbs in a large-text corpus. In effect, the action verbs, for which elicited neural activity was known, provided a grounding or bases for representing associated nouns.
These corpus models, such as Latent Semantic Analysis, use the co-occurrence of words within some context (e.g. a document) to learn lower dimensional, vector representations of word concepts (Deerwester et al. 1990). Like the reviewed psychological research (Jones and Love 2007), words need not directly co-occur with one another to become more similar, but need only occur in similar contexts. Although LSA has enjoyed numerous successes, cases in which its representations diverge with those of humans has prompted further model development (Wandmacher et al. 2008).
One subsequent proposal, Latent Dirichlet Allocation (LDA), is a probabilistic approach in which documents are generated according to a mixture of probabilities over latent themes or topics (Blei et al. 2003). For example, LDA may find that the words ‘Parliament’ and ‘Prime Minister’ have a high probability of belonging to the same topic (e.g. ‘politics’). A passage about the Prime Minister visiting the Houses of Parliament would make this politics topic highly probable, though other topics would also be somewhat likely, such as a topic related to tourism (Big Ben is part of the Houses of Parliament).
The representations learned by topic models appear similar to the concepts that people use (Griffiths et al. 2007; Andrews et al. 2009). For example, topic modelling can predict subsequent words in a sentence, disambiguate word meanings, and extract the gist of a sentence (Griffiths et al. 2007). Related techniques find that word meanings extracted for text corpora reflect back that society’s gender stereotypes (Bolukbasi et al. 2016). These successes emphasize the importance of extrinsic roles and relationships.
People learn thematic relations by observing co-occurence in events or situations (Estes et al. 2011). In corpus analysis, word co-occurrence in language is assumed to be a proxy for co-occurrence in the wild. However, this assumption may not always hold. For example, words can co-occur in language without being semantically related (e.g. iceburg→ lettuce). More generally, most spoken language is concerned with effective communication of relevant information (Grice 1975), rather than providing a faithful record of object interactions. For example, in waiting to cross the street with a companion, one would never verbalize that the passing car drives on the road. Written language also tends to be curated. For example, journalists adhere to particular guidelines and aim to report on stories of interest to their readership. Whether it’s from natural language or otherwise, data that captures co-occurrence of events in the wild is best suited to evaluate the structure of people’s thematic representations.
An alternative dataset that may help to further evaluate the influence of extrinsic features on people’s representations is consumer retail data. Retail data are collected from consumers as they purchase products together in the same basket, analogous to how words group together in the same document (see Fig. 1). Whilst a person may be conscious not to voice every item they bought in their supermarket shop, one’s grocery receipt provides a faithful record of what they purchased in a supermarket visit. Importantly, this data is traceable to an individual, which contrasts with most corpora analyses, which tends to be based on language in newspapers and books (e.g. Griffiths et al. 2007). Large-scale analyses of grocery retail data is therefore well placed to evaluate the claim that individual differences in people’s experience of the world leads them to possess different thematic representations. In particular, it may help to supplement existing research investigating how people cross-classify food (Murphy and Ross 1999; Ross and Murphy 1999; Lawson et al. 2017; Blake 2008), such as elucidating how regional and generational differences affect people’s thematic representations.
An additional benefit of using consumer purchasing data is that it suits the mathematical assumptions of topic models particularly well. For example, natural language researchers typically use their domain expertise to remove function or ‘stop’ words that have little semantic meaning (such as the, of, and). They may also ‘stem’ words to remove prefixes and suffixes of words that have similar semantic meaning (e.g. eat vs. eating). Moreover, the order of words in sentences can also make a big difference to sentence meaning (e.g. ‘dog bites man’ vs. ‘man bites dog’). However, most standard implementations of topic models (based on the original algorithm by Blei et al. 2003) typically ignore word order, instead preferring to consider language as a ‘bag-of-words’ (for an alternative, see Huang and Wu 2015). In contrast, for retail data captured in-store, there is no inherent order for products within a basket, nor a need to remove stop words or perform stemming.
If people’s thematic organization of concepts arises through their interaction with the environment, then it should be possible for a topic model to recover relevant representations of these through consumer purchasing patterns, as shown in Fig. 2. Whilst earlier research has indicated that this is possible, none (to this author’s knowledge) have explicitly measured the likeness of learned topics to consumer’s mental representations (Iwata and Sawada 2013; Iwata et al. 2009; Hruschka 2014). Although people have been shown to default to a taxonomic organization (e.g. tomato → fruit) when asked to freely sort food items in the lab, the presence of a goal can lead to a thematic organization (e.g. tomato → salad) during decision-making (Murphy and Ross 1999; Ross and Murphy 1999). Because shopping is highly goal-directed, we hypothesized that the topics recovered by a topic model would reflect thematic organization. We tested these predictions using a large, anonymized dataset of 1,252,963 shopping baskets and 5,753 unique products, suppliedby one of the UK’s largest supermarket retailers. After optimizing an LDA solution using fit statistics and checking for convergence,Footnote 1 we labelled the 25 topics recovered by the model.
To foreshadow, we found that LDA recovered meaningful topics that were primarily goal-directed and thematic in nature. We confirmed the psychological reality of these topics in two human studies, one with judgments from retail experts and another involving typical consumers. Further support came from analyses showing that topics tied to a season varied sensibly in their prevalence over the calendar year (e.g. the Christmas topic was most prevalent in December). Overall, these results suggest that—contrary to early research on cross-classification of food (Murphy and Ross 1999; Ross and Murphy 1999)—thematic relations dominate representations of food. This is in line with more recent claims that thematic relations may be more numerous than taxonomic associations in people’s stored semantic network and may be more easily revealed when examined at scale (Estes et al. 2011; Lawson et al. 2017). Final analyses tested whether an individual’s shopping experience shaped their conceptual organization of the products. In support of this assertion, the rate at which an individual sampled topics (based on recent shopping history) predicted the individual’s age, gender, and geographic region. This suggests that individual differences in people’s experience can lead them to possess notably different thematic representations from each other. This is important, because it suggests that food-related themes discussed in the literature are likely a function of the participants’ individual experiences and culture.