Word meaning in minds and machines

1 Introduction

• NLP systems have achieved strong performance across language tasks, but engineering success may not reflect the semantic processing used by human speakers and listeners.Their capabilities include translation, summarization, question answering, natural language inference, and practical customer-facing applications.Source: §1 · pp. 1–2
• The article evaluates whether NLP word representations can serve as models of psychological semantics, while distinguishing functional similarity from human-like learning mechanisms.The authors focus on whether derived word representations resemble those people acquire, not whether models learn meanings in the same way.Source: §1 · p. 2
• NLP need not prioritize psychological plausibility for every application, but more human-like word representations could address limits in language comprehension and production.Models intended as psychological theories must also account for how people use language in everyday life, beyond current benchmarks.Source: §1 · p. 2

1.1 Coverage of this article

• The article reviews psychological theories, model desiderata, and NLP approaches before arguing that current representations support word similarity but not the flexible semantic behaviors people rely on.It also considers whether psychologically realistic semantics could improve NLP on advanced understanding and communication problems.Source: §1.1 · pp. 2–3

Referential and possible-world semantics

• Referential semantics links words to objects or states of affairs, but simple reference is unstable because the relevant sets of entities constantly change over time.Formal semantics addresses this problem with extensions across time- and context-indexed possible worlds, including hypothetical and fictional situations.Source: §2 · p. 3

Meaning and world knowledge

• Word meaning cannot be cleanly separated from general knowledge: hearing a word provides access to knowledge needed for interpretation, while NLP evaluations often engage capabilities beyond isolated word meaning.The authors therefore avoid committing to a precise boundary between linguistic meaning and world knowledge.Source: §2 · pp. 4–5

Conceptual semantics

• Psychological theories instead treat word meaning as access to conceptual structure, allowing words such as dog or car to activate perceptual, functional, and inferential knowledge.This account explains how listeners infer unstated information during comprehension while avoiding representations of every object in the world or every possible world.Source: §2 · pp. 4–5

Grounding in perception and action

• The conceptual approach connects language to the world through perceptual and motor mechanisms, making word use causally related to objects and enabling language to guide thought and action.It is psychologically plausible because people possess concepts and world knowledge, unlike infinite sets of objects in possible worlds.Source: §2 · pp. 3–5

Learning and scope

• Word learning can reshape conceptual structure, as learning labels such as goat encourages children to distinguish objects that might otherwise remain within a broader category.The authors focus on word representations tested in contextual use, while excluding sentence- and discourse-level semantic phenomena from the investigation.Source: §2 · p. 6

Five desiderata

• A psychological-semantics model must support five basic language functions, including flexible physical, verbal, and mental behaviors grounded in conceptual representations.The desiderata are intended as minimum requirements: a model failing one or more would fall far short, although satisfying them would not guarantee fully human-like language.Source: §3 · p. 6, Table 1 · p. 7
• Words must connect perceptual input and mental representations to speech, listeners’ representations, and actions in the world.In the dinner-table example, describing a dirty knife requires recognizing objects and relations, while understanding the request enables another person to replace the unsuitable knife.Source: §3 · pp. 6–7
• Language must express goals and support appropriate action, as when “butter knife” communicates the desired kind of object rather than a generic knife.The speaker forms a goal, translates it into a phrase, and the listener identifies the intended object instead of a spoon, plate, or other alternative.Source: §3 · p. 7
• Word representations must support flexible conceptual combination, allowing speakers to construct and listeners to understand novel phrases such as “butter knife.”Understanding the phrase requires combining concepts to infer a shorter, blunt knife made for butter or cheese.Source: §3 · p. 7
• Language both requires world knowledge for comprehension and serves as a medium for expanding knowledge that can change future behavior.General statements such as “Forks go on the left, knives on the right” communicate information beyond the immediate table and may guide later table-setting actions.Source: §3 · p. 7

4 Computational Approaches to Word Meaning

• The section therefore distinguishes computational meanings derived from text relations from psychological semantics, so claims about what has been solved are not conflated across fields.A psychological theory of semantics must meet the requirements outlined earlier in Table 1.Source: §4 · p. 8

4.1 Word Representations

• Distributional models represent word meaning through usage patterns, with LSA embeddings capturing co-occurrence relations and supporting similarity judgments, lexical tasks, and limited composition.LSA reduces a word-by-document co-occurrence matrix to vectors, while predictive models such as CBOW learn embeddings by predicting words from context.Source: §4.1 · pp. 8–10

4.2 Sentence representations

• Sentence meaning cannot be captured by simply adding word vectors, because this loses word order and the relations that distinguish different propositions.Language models, including RNNs and Transformers, use contextual processing and word order to learn syntactic and semantic structure, while Transformers resolve meanings such as river versus financial bank from context.Source: §4.2 · pp. 10–12

4.3 NLP as a theory of semantics

• Although large neural networks provide strong performance in translation, summarization, question answering, and inference, their success does not by itself establish a theory of psychological semantics.NLP is primarily driven by application-specific performance, whereas psychologists have tested model agreement with human data as evidence for psychological validity.Source: §4.3 · pp. 12–13

Osgood’s theory

• Osgood’s factor analysis reduced adjectival ratings to evaluative, potency, and activity dimensions, whose values predicted some behavioral similarities and responded sensibly to experimental manipulations.The approach operationalized semantics behaviorally rather than through mental representations.Source: §5 · p. 13

Similarity scaling

• Low-dimensional scaling solutions predicted categorization and sentence-evaluation performance by representing the overall similarity among items.Human pairwise similarity ratings were reduced into spatial representations, where distances such as bear–mammal predicted classification times.Source: §5 · p. 13

Limitations

• Scaling solutions and Osgood’s dimensions represented similarity but lacked the specific knowledge required to identify concepts and use words in normal language.Animal identification can require perceptual forms, bodily parts, behavior, habitats, and other facts absent from low-dimensional representations.Source: §5 · pp. 13–14
• These early models could not describe perceptual scenarios, explain listeners’ understanding, or select words according to desires and goals, although historical computing and technical constraints limited what researchers could build.The limitation concerns the primary desiderata of a semantic theory, not merely the dimensionality of the representations.Source: §5 · p. 14

Transition to NLP

• Human-judgment scaling was computationally laborious, whereas LSA and later NLP models used large-corpus word co-occurrence patterns to acquire semantic information more efficiently.For N items, the similarity matrix requires N*(N-1)/2 entries, each averaging multiple human ratings.Source: §5 · p. 14

6 Semantic similarity

• LSA mishandles inflectional similarity, rating computed only .35 to compute and saddle more similar to horse than horses, while newer models better capture these variants.The newer models still make semantic-detail errors, such as ranking saddle above other mammals as a neighbor of horse.Source: §6 · pp. 15, 17
• CBOW’s nine nearest neighbors for “dog” include morphological variants, subordinate breeds, a superordinate, and related mammals, avoiding LSA’s cross-domain actions and objects.The reported similarities range from dogs (0.85) to cat (0.71), although corpus size and model selection may affect the result.Source: §6 · pp. 15–16
• BERT and GPT-2 produce semantically coherent “dog” neighbors from the same domain, but their rankings still misrepresent some fine-grained similarities and omit expected canines.BERT places horse ahead of some closer canine terms, while GPT-2 is case sensitive and also places horse before other canines.Source: §6 · p. 16
• BERT shifts “bank” toward financial or aquatic neighbors according to sentence context, whereas GPT-2 behaves similarly for bank but fails to switch lead interpretations across contexts.BERT’s contextual embedding favors treasury and ATM in financial contexts and shore and beach in aquatic contexts; GPT-2 favors advantage in all tested lead contexts.Source: §6 · pp. 17–18, Table 2 · p. 18, Table 3 · p. 18
• Large-scale predictive models improve on LSA by capturing same-domain semantic neighbors and resolving word meanings from context, though important details remain incorrect.The authors argue that CBOW, BERT, and GPT-2 are preferable replacements for LSA in psychological studies, while noting that models remain imperfect.Source: §6 · pp. 17–19

6.1 Human Behavioral Judgments and Ratings

• Human-rating comparisons do not necessarily measure word meaning because association and priming are strongly influenced by frequency, contiguity, and co-occurrence.Lucas (2000) found semantic similarity produced about half the priming effect of association, so embedding cosines that predict both constructs may conflate them.Source: §6.1 · p. 19

7 Desiderata

• Text-only NLP systems represent words through other words and therefore cannot connect linguistic meaning to perception, action, or world entities without suitable interfaces.A model need not itself interact with the world, but it must contain representations capable of describing objects, properties, relations, and categories through appropriate interfaces.Source: §7.1 · p. 20, §7 · p. 20

7.1 Word representations should support describing a perceptually present scenario, or understanding such a description.

• Human word meaning requires coherent knowledge of perceptual features, parts, relations, uses, and implications so speakers can understand sentences and draw appropriate inferences.For example, understanding a knife requires linking its blade and handle to functions, danger, repair, and alternative tools that can fulfill its role.Source: §7.1 · pp. 21, 27, 28
• Vision-language systems use visual and language encoders to align images with words, enabling captioning, question answering, instruction following, and some zero-shot category prediction.Captioning architectures may pass localized visual embeddings to a language decoder, which can attend to relevant image regions while generating words.Source: §7.1 · pp. 21–23
• Captioning and recognition systems overgeneralize from superficial visual features: they mistake cars or dressers for suitcases and fail to recognize natural umbrellas or describe them accurately.For natural umbrellas, ResNet50 ranked “umbrella” at a median position of 181 out of 1000, with a highest rank of 42.Source: §7.1 · pp. 25–27
• Multimodal models ground words in perception but still fall short of psychological semantics because they lack structured conceptual knowledge about objects, relations, functions, and goals.They can describe some perceptual scenarios and answer certain questions, but neuro-symbolic representations may be needed to bridge modalities and support richer semantic content.Source: §7.1 · pp. 21, 23, 28

7.2 Word representations should support choosing words on the basis of internal desires, goals, or plans.

• A theory of lexical semantics must explain how speakers translate internal thoughts, desires, goals, and plans into words, not merely label externally present objects.Text-based systems lack nonlinguistic representations of ideas, while goal-directed dialogue systems connect language to entities and actions only within narrow application domains.Source: §7.2 · pp. 28–29
• Dialogue systems rarely choose words from explicit desires, goals, or plans: goal-directed systems use constrained or hand-coded worlds, while chit-chat models react statistically and lack grounding.These systems cannot flexibly name novel objects, form novel combinations, learn new facts verbally, or develop reliable internal goals from text alone, though persona and grounded-dialogue efforts are emerging.Source: §7.2 · pp. 29–30

7.3 Word representations should support responding to instructions appropriately.

• Instruction-following models generalize some object, color, action, and physical-property relations, but fail catastrophically on novel combinations and abstract meanings such as not, move, size, and adverbs.Generalization includes new object-color pairs, actions, and heavy-object relations; however, southwest movement reaches 0% correct, and not generalization falls below 50% in a 2D world and 60–80% in a 3D environment.Source: §7.3 · pp. 31–32

7.5 Word representations should support changing one’s beliefs about the world based on linguistic input.

• BERT’s taxonomic predictions change sharply with minor wording changes, suggesting it has partial lexical associations but lacks stable, general knowledge that can be reliably updated through language.For example, BERT predicts mammals over birds and fish after “but,” birds after “and” or “while,” and fish for “A dolphin is a”; dolphin’s cosine similarities also vary by inflection.Source: §7.5 · pp. 36–38
• BERT’s animal-category accuracy was 54.8% singular and 77.4% plural, while GPT-2 reached 83.9% and 77.4%, yet singular–plural answers mismatched 51.6% and 35.5%, respectively.These inconsistencies indicate that performance does not reflect stable taxonomic beliefs.Source: §7.5 · p. 38
• GPT-2 generated fluent unicorn text but contradicted itself immediately by describing unicorns as both one-horned and four-horned, suggesting that correcting a belief would be difficult.Its 1.5 billion parameters make it unclear how a specific fact could be opened up and revised.Source: §7.5 · p. 39
• Factoid accuracy remained low: GPT-2 answered 4.1% of Natural Questions correctly, whereas GPT-3 reached 15-30%, and retrieval techniques substantially improved performance.BERT can make reasonable ConceptNet predictions, but these remain brittle and insufficient for conceptual combination and broader language understanding.Source: §7.5 · p. 39
• Semantic representations must encode labeled relations such as ISA and CAN, because unlabeled links between words are insufficient for coherent inference.Current NLP models rely purely on text predictability rather than learning explicit relations.Source: §7.5 · pp. 39–40

7.6 Summary of Desiderata

• Successful systems should ground implicit relations in perception, extract them from text, use knowledge bases for consistency, and revise beliefs from linguistic input.Multimodal grounding in vision and action and more abstract semantic representations are proposed to connect language with world knowledge, action, and knowledge.Source: §7.5 · p. 40, §7.6 · p. 40

8 Successes in NLP

• Modern NLP systems learn syntax, semantics, word associations, and some world knowledge from massive textual patterns, making them useful when words are both inputs and outputs.Pretraining, architectural innovations, large datasets, computing resources, and fine-tuning amplify these capabilities.Source: §8 · pp. 40–41
• Distributional models can approximate relations such as capitals, inflections, plurals, and some syntactic dependencies, but text alone does not represent what words refer to.For example, similar contexts can link knife and knives without representing singularity, plurality, or what a knife is.Source: §8 · p. 41

8.2 Past Critiques Within Psychology

• Human readers often supply the world knowledge and inference that make model outputs meaningful, whereas practical psychological semantics requires connecting words to concepts and the outside world.Psychological critiques focused only on word-pair relatedness cannot test generating descriptions from perception or changing world knowledge through instructions.Source: §8 · pp. 41–42, §8.2 · p. 45

8 Successes in NLP

• Future NLP progress must move beyond pattern recognition toward reasoning, compositionality, concept learning, multimodal grounding, background knowledge, abstraction, goals, and beliefs.The authors identify neuro-symbolic modeling and grounded language learning as promising directions for more psychologically realistic semantics.Source: §8 · p. 42

8.1 GPT-3 and Scaling Up

• GPT-3’s 175 billion parameters and approximately 500 billion training tokens—25 times GPT-2’s data—did not add mechanisms for grounding, goals, complex novel compositions, or belief change.It remains trained from text alone, predicts the next word rather than expressing internal goals, and frequently contradicts itself.Source: §8.1 · pp. 43–44
• GPT-3 is a strong few-shot learner, reusing text-trained representations across novel tasks without retraining, although task-specific models usually perform better.It often continues question answering, grammar correction, or numerical addition after receiving only a few examples in a prompt.Source: §8.1 · p. 43
• 47% and 58% on RACE, versus about 90% and 93% for the best NLP model, show GPT-3 can generate coherent continuations without fully understanding passages.GPT-3 also performs poorly on adversarial benchmarks that separate surface patterns from semantic content, suggesting limits to scaling the pretraining objective.Source: §8.1 · p. 44

8.2 Past Critiques Within Psychology

• Earlier psychological studies tested whether perceptual, affective, or textual information improves fits to human judgments, but they did not address the broader goals of semantic theories.These evaluations mainly remained within relatedness measures for words or sentences, while embodiment is only one version of the more general world-connection problem.Source: §8.2 · p. 45

9 Conclusion

• Text-based NLP models will continue to advance and perform useful tasks, but they alone will not form the basis of a psychological theory of word meaning.The authors distinguish practical task performance from modeling psychological semantics.Source: §9 · p. 46
• For task-specific applications with abundant labeled data, large-scale pattern recognition or fine-tuning may be sufficient, so psychological semantics is not always the goal.The paper’s arguments have little relevance to cases such as domain-specific question answering with many thousands of training pairs.Source: §9 · p. 46
• Flexible language understanding, responses to novel input, and genuine conversation may require a conceptually based compositional semantics closer to human semantics.The authors present psychological semantics as a higher bar that could also improve future language applications.Source: §9 · p. 46
• Semantic representations should connect words to content and the world, rather than relying only on word-to-word patterns, and should encode internal structure among concepts and their constituents.Knowing that dogs relate to legs, fur, toys, and mammals is insufficient without representing the specific relations among those concepts.Source: §9 · p. 46
• Human word learning is grounded partly in perceptual-motor experience, which supplies information about objects and their properties that verbal descriptions may not adequately communicate.Children can learn a lion’s parts, shape, color, sounds, and behaviors through perception, motivating hybrid visual-language models.Source: §9 · pp. 46–47

9.1 Final Conclusion

• Complete models of human word meaning require conceptual structures linked to words and built through perception, action, reasoning, and interpreted language, not textual statistics or word prediction alone.Multimodal models are moving toward these goals, but their linguistic abilities and representations remain too limited and pattern-bound to establish human-equivalent meanings.Source: §9 · p. 47, §9.1 · p. 48

• Authorship is alphabetical, and the authors thank eight colleagues for comments on an early draft.Source: Authors’ Note · p. 48
• B. Lake’s contribution was partially funded by NSF Award 1922658 and DARPA Award A000011479; PO: P000085618.Source: Authors’ Note · p. 48
• A previous version of the manuscript appeared on arXiv.org as 2008.01766.Source: Authors’ Note · p. 48