Word meaning in minds and machines

Brenden M. Lake; Gregory L. Murphy

Source-linked AI summary

Word meaning in minds and machines

Brenden M. Lake, Gregory L. Murphy

arXiv:2008.01766v3Aug 4, 2020cs.CLcs.AIcs.LG

TL;DR

The paper asks whether current NLP models capture human word meaning beyond their success on linguistic tasks and word similarity. It compares machine and human representations, arguing that more human-like semantics must connect words to perception, action, goals, beliefs, and flexible conceptual combination.

Problem
The central question is whether NLP systems that succeed at language tasks also provide psychologically adequate models of human word meaning.
Method
The paper compares contemporary NLP representations with human semantic requirements, including similarity, grounding, internal states, beliefs, and compositional meaning.
Results
Current NLP systems model human word similarity fairly well but fall short on grounding meaning in perception and action, internal goals and beliefs, and flexible conceptual combination.
Takeaways & Limitations
More successful language systems and psychological theories may require conceptually based compositional semantics connected to perceptual experience and world knowledge.
Takeaways & Limitations
Multimodal models have not yet reached human performance, likely because they lack conceptual knowledge for interpreting visual training data.

Abstract

from arXiv · show

Machines have achieved a broad and growing set of linguistic competencies, thanks to recent progress in Natural Language Processing (NLP). Psychologists have shown increasing interest in such models, comparing their output to psychological judgments such as similarity, association, priming, and comprehension, raising the question of whether the models could serve as psychological theories. In this article, we compare how humans and machines represent the meaning of words. We argue that contemporary NLP systems are fairly successful models of human word similarity, but they fall short in many other respects. Current models are too strongly linked to the text-based patterns in large corpora, and too weakly linked to the desires, goals, and beliefs that people express through words. Word meanings must also be grounded in perception and action and be capable of flexible combinations in ways that current systems are not. We discuss more promising approaches to grounding NLP systems and argue that they will be more successful with a more human-like, conceptual basis for word meaning.

1 Introduction

The article asks whether NLP word representations can model human psychological semantics. It argues that current systems succeed on some language tasks and word similarity but require broader, more human-like representations for flexible language use.

NLP systems have made major progress on translation, summarization, question answering, and natural language inference.
Psychologists have proposed that NLP-derived word representations could be functionally similar to human representations.
The article evaluates whether NLP systems can represent human semantic knowledge, focusing on word meaning rather than the stronger claim that humans learn meanings like machines.
Flexible human language behavior requires representations linked to world knowledge, perception, goals, and everyday language use beyond current benchmarks.
Current representations are useful for modeling word similarity, but their details do not always align with human semantic similarity.

2 Semantics in Cognitive Science

Psychological semantics treats word meanings as conceptual structures connected to world knowledge and the world, rather than as merely changing sets of referents. Language both uses this knowledge and changes it.

Referential semantics analyzes meaning as a relationship between words and worldly entities or situations, including hypothetical possible worlds.
Language comprehension depends on world knowledge, as listeners infer unstated consequences from statements such as “Fred is my dog.”
A central challenge is deciding which features belong to word meaning and which belong to general world knowledge.
Word meanings are embedded in world knowledge, and linguistic input can change people’s concepts and knowledge.
Psychological semantics predominantly treats words as pointers to conceptual structures that organize world knowledge.
Conceptual representations connect language to the world through perceptual and motor mechanisms while remaining psychologically plausible.
The article limits its investigation to word representations in context, excluding sentence- and discourse-level semantic phenomena such as entailment and ambiguity.

3 Desiderata for a model of psychological semantics

The authors propose five desiderata for psychological-semantic representations, centered on supporting flexible physical, verbal, and mental behaviors. These functions connect language with perception, knowledge, and action.

An adequate theory must support basic language-use behaviors, including flexible physical, verbal, and mental activities grounded in conceptual representations.
Producing a description requires recognizing perceptually present objects and relations, then retrieving words that refer to them.
Understanding language can guide action, as a listener identifies the intended object and replaces an unsuitable knife.
Words can be activated by perception and mental representations, influence listeners’ representations, and support novel conceptual combinations such as “butter knife.”
Language both requires knowledge for comprehension and serves as a medium for increasing knowledge.

4 Computational Approaches to Word Meaning

Computational approaches derive word representations primarily from text co-occurrence or word-prediction objectives, progressing from LSA and CBOW to contextual language models. These methods support similarity and prediction but struggle with richer sentence relations and psychological semantics.

4.1 Word Representations: Distributional semantics defines similar word meanings through similar usage and co-occurrence patterns.
4.1 Word Representations: LSA represents word meaning with vectors derived from a word-by-document matrix reduced by singular value decomposition.
4.1 Word Representations: LSA measures similarity with vector cosine or dot products and supports tasks such as synonym selection, sentence completion, summarization judgments, and part of the TOEFL.
4.1 Word Representations: Text-co-occurrence models are limited in identifying semantic relations beyond similarity, motivating methods augmented with labels or other structure.
4.1 Word Representations: Predictive models learn representations by predicting word probabilities from context, with CBOW averaging context embeddings before comparing candidate outputs.
4.1 Word Representations: Figure 1 contrasts CBOW, RNN, and BERT architectures, distinguishing word embeddings from context-sensitive hidden embeddings and showing their prediction objectives.
4.2 Sentence Representations: Summing word embeddings cannot capture the full richness of sentence meaning because comprehension requires recovering relations and modeling syntax and sentential semantics.
4.3 NLP as a theory of semantics: NLP is primarily driven by engineering goals and task performance, while claims that these models constitute theories of psychological semantics come mainly from psychologists.

5 Early Computational Theories of Psychological Semantics

Early psychological theories represented semantic similarity through ratings and low-dimensional spaces, but these representations omitted the detailed knowledge people need to use words flexibly.

5 Early Computational Theories of Psychological Semantics: Psychological scaling represented similarity among words or concepts and helped predict comparison-based behaviors such as classification and memory confusions.Multidimensional scaling and clustering made similarity structures comprehensible, including mammal and bird concepts.
5 Early Computational Theories of Psychological Semantics: Similarity dimensions such as animal size and predacity distinguished stimuli globally but could not identify particular mammals reliably.The dimensions do not specify which animal is, for example, midsized and moderately predacious.
5 Early Computational Theories of Psychological Semantics: Using names such as sheep, goat, or cow requires perceptual, biological, and behavioral knowledge absent from Osgood’s evaluative dimensions.Knowing evaluation, potency, and activity does not reveal grammatical category, concreteness, or semantic domain.
5 Early Computational Theories of Psychological Semantics: Low-dimensional scaling spaces omit the knowledge that supports identifying specific items, describing scenarios, understanding descriptions, and choosing words for goals.Distinguishing sheep from goats may require detailed bodily shapes, proportions, and parts.
5 Early Computational Theories of Psychological Semantics: Human-rated similarity matrices become laborious at scale because N items require N*(N-1)/2 pairwise entries, motivating corpus-based alternatives such as LSA.The entries are averages of multiple human ratings.

6 Semantic similarity

NLP embeddings increasingly capture semantic similarity and contextual shifts, outperforming older LSA representations in several respects, but remain limited by corpus-based relations and incomplete meaning representations.

6 Semantic similarity: LSA separates morphological variants and can assign low similarity to synonymous tense variants, illustrating how text-based representations may diverge from semantic similarity.Compute and computed receive cosine similarity .35, while saddle is more similar to horse than horses is (.91 vs. .83).
6 Semantic similarity: Modern predictive models produce semantically coherent nearest neighbors, whereas LSA often mixes similarity with actions, associations, and subordinate relations.BERT’s neighbors for dog remain within the same semantic domain, though their rankings are imperfect.
6 Semantic similarity: More powerful models capture inflectional and morphological similarity, finding pairs such as horse–horses and compute–computed close together.These patterns address shortcomings found in earlier count models such as LSA.
6 Semantic similarity: Contextual BERT embeddings shift homonym representations toward the relevant sense: bank favors financial meanings ambiguously but water meanings in aquatic contexts.Targeted comparisons use treasury, ATM, shore, and beach under matched sentence frames.
6 Semantic similarity: BERT similarly resolves lead by context, favoring advantage in ambiguous or sports contexts and metal-related meanings when physical substance is evoked.GPT-2 instead preferred the advantage interpretation across the tested contexts.
6 Semantic similarity: These homonym tests show contextual movement toward a correct meaning, not detailed representations of its properties or rejection of other senses.A complete model would represent properties such as weight, flexibility, toxicity, and radiation blocking for lead.
6 Semantic similarity: Word association and priming are not direct measures of word meaning because frequency and co-occurrence can connect related but semantically dissimilar words.Semantic similarity contributes less to priming than association, and relations such as cat–bowl do not specify what a cat is.

7 Desiderata

Adequate word representations must connect language to perception, action, goals, and conceptual knowledge rather than only textual patterns. Current multimodal and goal-directed systems make progress, but remain brittle, insufficiently structured, and tied to training experience.

7.1 Word representations should support describing a perceptually present scenario, or understanding such a description.: Word representations should connect language to perceptual input and support understanding and describing present scenarios.Multimodal architectures combine visual and language encoders, but current systems still lack conceptual knowledge for interpreting images.
7.1 Word representations should support describing a perceptually present scenario, or understanding such a description.: Vision-language models overgeneralize from superficial visual similarities, producing inaccurate captions for novel or abstract scenes.Captioning systems confuse cars or dressers with luggage and struggle with natural umbrellas, whose unusual object-animal combinations disrupt identification.
7.2 Word representations should support producing words based on internal states.: Word representations should support internal goals and plans, yet current goal-directed systems often generate canned text without graded word representations.Human language production reflects absent objects, errors, unusual situations, desired future states, and facts intended for interlocutors.
7.4 Word representations should support producing and understanding novel conceptual combinations.: Word representations should support structured conceptual knowledge and novel combinations, but current models lack robust world knowledge and independently learned structure.Text-based learning has not yielded the structured knowledge available in hand-designed systems, while multimodal grounding alone does not encode object parts, functions, and relations.
7.3 Word representations should support responding to instructions appropriately.: Word representations should support responding to instructions through flexible, composable meanings that generalize beyond practiced objects, colors, and actions.Instruction-following models can generalize some object and color combinations, but struggle with negation and other abstract meanings.

8 Successes in NLP

Modern NLP systems learn extensive linguistic patterns from massive text corpora and perform impressively across many language tasks. However, their success remains limited as a model of human meaning because text-based representations lack grounding in the world, goals, beliefs, and robust semantic reasoning.

8 Successes in NLP: Text-trained models predict human similarity judgments because massive corpora provide word associations and patterns useful when words are both inputs and outputs.These systems can learn regularities such as related words, inflections, and plurals without explicitly representing the underlying concepts.
8 Successes in NLP: Relations among textual entities are useful for practical systems, and people can often supply semantic context when interpreting their outputs.Applications include information retrieval and restaurant reservations, while text-based models remain unable to interactively use language in the world as humans do.
8.1 GPT-3 and Scaling Up: GPT-3 performs many tasks from text prompts without task-specific fine-tuning, although it usually trails specialized models.Its few-shot flexibility resembles the reuse of human semantic representations for new tasks.
8.1 GPT-3 and Scaling Up: GPT-3 still lacks mechanisms for grounding words in beliefs and can produce internally contradictory descriptions despite impressive language performance.Its text-only training and next-word objective leave it vulnerable to contradictory corpus information and incoherent claims.
8.1 GPT-3 and Scaling Up: GPT-3 performs poorly on adversarial benchmarks that separate surface patterns from underlying semantic content, suggesting limits to scaling the pretraining objective.The proposed research frontier therefore emphasizes reasoning, compositionality, concept learning, multimodal learning, neuro-symbolic modeling, and grounded language learning.
8 Successes in NLP: Current NLP systems remain inadequate psychological theories because their word representations are not connected to perception, action, or the external world.Text-only vectors cannot support labeling objects or describing scenes from visual features, and words remain meaningful mainly through relations to other words.

9 Conclusion

The authors argue that text-based NLP models can achieve useful language performance but cannot alone explain human word meaning. More human-like systems should ground words in concepts, perception, action, and structured knowledge of the world.

9 Conclusion: Text-based NLP models may achieve important language capabilities, but they alone will not form a psychological theory of word meaning.The authors distinguish engineering success on particular tasks from psychologically adequate semantic representations.
9 Conclusion: A psychologically motivated semantics may improve productive, flexible understanding, responses to novel input, and actual conversations.The authors connect these abilities to conceptually based compositional semantics.
9 Conclusion: Semantic representations should connect language to content that makes contact with the world rather than only linking words through textual patterns.The authors argue that word relations alone do not support talking about actual things or obtaining information from language.
9 Conclusion: Word meanings require internal structure that represents categories, parts, and the relations among concepts.Knowing that dogs relate to leashes or fur is insufficient without representing relations such as category membership and part-whole structure.
9 Conclusion: Human word learning is supported by perception, action, and world knowledge, which let learners acquire properties and relations that language alone may not communicate.The authors use children learning about lions through perception as an example and identify hybrid visual-language models as a promising but difficult direction.
9 Conclusion: Multimodal models are moving toward grounded semantics, but their linguistic abilities remain limited and tied to training patterns, while complete human-like meaning requires a conceptual world structure.Text can contribute to learning such a structure, but systems would need to interpret relations, categories, and properties rather than merely predict words.

Authors’ Note

The authors report alphabetical authorship order, acknowledge contributors to an early draft, and identify funding supporting B. Lake’s contribution.

Authors’ Note: The paper lists its authors in alphabetical order.
Authors’ Note: The authors thank eight colleagues for comments on an early draft.
Authors’ Note: B. Lake’s contribution was partially funded by NSF and DARPA awards supporting data science and machine common sense research.