PaperlayerView exact revision on arXiv

Source-linked AI summary

A Primer on Computational Semantics for Artificial Intelligence Systems

Casey Kennington

arXiv:2608.25022v1cs.CLcs.AI

TL;DR

The paper addresses how artificial intelligence systems learn and represent linguistic meaning, and why language itself must be understood before meaning can be conveyed to machines. It surveys scientific, philosophical, and computational perspectives, including formal, grounded, and distributional semantics. It concludes that text-trained language models can understand language abstractly but lack embodied and emotional experience for deeper meanings, motivating a curriculum closer to human learning and a unified semantic approach.

  • Problem

    The paper asks how language models learn and represent meaning, given that text is not identical to language or meaning and human meaning depends on experience.

  • Method

    The paper synthesizes scientific and philosophical perspectives on language, explains formal, grounded, and distributional semantics, and compares transformer models with human language learning.

  • Results

    Transformer-based distributional models improved NLP applications and benchmarks, while text-trained models can understand language abstractly but lack embodied and emotional experience for deeper meanings.

  • Takeaways & Limitations

    The paper supports computational semantics that follows a curriculum similar to human learning and unifies distributional, grounded, and formal semantics.

  • Takeaways & Limitations

    Models trained only on text lack a notion of concrete word meanings, and text-based affordance knowledge reflects what has been written rather than established model meaningfulness.

Abstract

from arXiv · show

As people adopt transformer-based language models (e.g., ChatGPT and Gemini) for an increasing number of use-cases, it is important to know how such models learn and represent the meaning of the language, and to be more informed about what language is. This document is an attempt to help the reader understand how linguistic meaning (i.e., semantics) is approached from different fields of scientific and philosophical examination. I also explain three primary semantic theories: formal semantics, grounded semantics, and distributional semantics then compare how transformer-based language models differ from how humans learn language.

1 Introduction

The paper begins with the difficulty of explaining familiar word meanings, using “red” to show that concrete categories are learned through experience. It argues that understanding meaning is necessary for responsibly conveying meaning to artificial intelligence systems.

  • The word red refers to a concrete category in the world, whose meaning people learn through experience rather than through a technical definition.Definitions describe connotations, whereas red is understood by identifying experienced examples.
  • People use language effectively without being able to explain the mechanics underlying their linguistic understanding.The paper compares this implicit linguistic knowledge with driving safely without understanding car engines.
  • The paper focuses on meaning because conveying meaning to machines requires understanding what meaning is.
  • It surveys language from multiple scientific and philosophical perspectives before examining formal, grounded, and distributional semantics.The planned discussion culminates in transformer language models and comparisons with human language learning.

2 What is Natural Language?

Natural language is symbolic communication that develops through community use, repetition, and change across spoken, signed, and written modalities. The section presents language as a multidimensional object studied by linguistics, semiotics, cognitive science, and neuroscience, while noting that the neural basis of meaning remains unclear.

  • Natural language is symbolic communication occurring within human communities through use, repetition, and change.Words symbolize categories or information without requiring the listener to directly perceive the referenced object.
  • Languages do not translate cleanly because individual words can encode distinctions that another language expresses differently.French si can affirm a positive interpretation of a negative-polarity question, while English yes may remain ambiguous.
  • Language is a process rather than a static object, and communication includes spoken, signed, and written forms.Speech and signing convey information through signals unfolding over time, while writing is only one communicative medium.
  • Linguistics studies language descriptively across phonetics, morphology, syntax, semantics, and pragmatics, whose boundaries can overlap.Semantics concerns meanings from words and subwords through phrases, sentences, and broader language use.
  • Semiotics examines communicative signals broadly, including words, gestures, smoke signals, and dances, together with their interpreters and interpretants.
  • Neuroscience investigates the brain and nervous system, but how meaning is stored and retrieved in the brain remains unclear.The paper identifies this as an area requiring substantial further work.

3 Philosophy of Meaning3

The philosophy section distinguishes reference from meaning and uses sense, connotation, denotation, language games, and compositionality to explain how linguistic meaning is structured and used. It concludes that composing meaningful larger expressions remains a challenge for computers.

  • Reference is the use of words to refer to objects, events, people, or entities within language itself.
  • Frege distinguishes sense, the meaning or notion of an expression, from reference, the entity to which it refers.Different expressions can share a reference while having distinct senses.
  • Mill distinguishes connotation, the properties implied by an expression, from denotation, what the expression applies to in the world.The dog example contrasts an abstract type of entity with a particular dog being denoted.
  • Wittgenstein’s language games frame language as use in context rather than a static set of facts.Examples include phone conversations and interactions with bookstore employees.
  • Compositionality holds that complex expression meanings are determined by the structure and meanings of their constituents.The principle applies across words, phrases, sentences, paragraphs, and dialogues.
  • Grammatical structure alone does not guarantee meaningful composition, making the construction of larger meanings a challenge for computers.“Colorless green ideas sleep furiously” illustrates a grammatical but nonsensical sentence.

4 Formal Semantics

Formal semantics represents linguistic meaning using logical machinery, translating language into structured expressions that can be evaluated as true or false. Its central limitation is that logic alone does not determine what real-world properties such as gray mean.

  • 4 Formal Semantics: Formal semantics uses logic as a computational basis for representing and processing linguistic meaning.The approach seeks to map spoken or written language into logical representations.
  • 4.1 Computation and Logic: Computational logic combines True/False values through operators such as AND, OR, and NOT to produce results.These simple operations can be connected to support complex computation.
  • 4.2 First Order Logic: First Order Logic provides operators and structures intended to represent relations between linguistic forms and meanings.
  • 4.2 First Order Logic: Formal semantic analysis translates statements into logical representations, such as combining properties with AND or expressing alternatives with OR.Examples include “big AND gray AND elephant” and “taco OR salad.”
  • 4.2 First Order Logic: Predicates can be treated as functions that return True or False for an assigned variable, while existential notation asserts that a suitable entity exists.
  • 4.2 First Order Logic: Combining predicates requires one variable to satisfy all specified properties simultaneously for the formula to be True.The big-gray-elephant example requires all three conditions at once.
  • 4.3 Shortcomings with Logical Forms: Logical forms presuppose an account of real-world properties and entities, involving set theory and modal logics.
  • 4.3 Shortcomings with Logical Forms: FOL does not solve how a machine determines what properties such as gray mean or when gray should return True.The section introduces grounded semantics as an attempt to address this shortcoming.

5 Grounded Semantics5

Grounded semantics challenges the idea that symbols or text carry meaning without connection to experience. It explains meaning through multimodal interaction with the world, while distinguishing concrete meanings that require grounding from abstract meanings often defined linguistically.

  • Symbol Grounding Problem: The Symbol Grounding Problem argues that symbols and text remain ungrounded unless connected to what they represent.The Chinese Room illustrates how manipulating symbols according to rules can produce outputs without establishing meaning.
  • Grounding in Experience: Grounded meaning arises from experience with the world, including visual perception, bodily interaction, sensory input, and muscle memory.The meaning of chair includes seeing chairs, using them for sitting, and feeling the bodily relief of sitting down.
  • Concrete and Abstract Meaning: Concrete words denote physical things and require Symbol Grounding, whereas abstract words denote ideas often defined by other words.Examples include chair and red as concrete words, versus democracy and travel as abstract words.
  • Concrete and Abstract Meaning: Concrete and abstract concepts form a continuum: visual depictions may clearly represent concrete words but only partially represent abstract ones.A depiction of voting captures one aspect of democracy rather than its full meaning.
  • Learning Meaning: Grounded semantics proposes bottom-up learning in which concrete words learned through interaction bootstrap more abstract categories, alongside top-down processing.Learning red, green, and yellow can support abstraction of the category color.
  • Shortcomings of Grounded Theories: Machine grounding remains unclear because multimodal representations may be combined with language models according to technical constraints rather than semantic considerations.Vision is the most represented grounding modality, although olfactory, tactile, haptic, vestibular, interoceptive, and muscular modalities also exist.

6 Distributional Semantics

Distributional semantics estimates word meaning from patterns of use in language, representing words as vectors whose relationships encode semantic similarity. This approach underlies word2vec and transformer language models, but text-only models lack concrete, embodied meaning.

  • Core idea: Distributional semantics estimates word meaning from how words are used within language, especially written text that computers can process efficiently.The approach is motivated by the intuition that words acquire meaning from the company they keep.
  • Vector representations: One-hot vectors assign each vocabulary word a unique position, but all words remain equally distant and therefore encode no semantic similarity.These vectors are large and sparse, motivating smaller, information-rich embeddings.
  • Vector representations: Context-counting methods build vectors from nearby words, so terms such as apricot and pineapple receive different profiles from digital and information.The resulting vectors capture recurring contextual associations rather than merely assigning arbitrary word numbers.
  • Semantic interpretation: Words with similar vector embeddings are treated as having similar meanings, with distance correlating with semantic similarity or difference.This provides a computational semantic theory based on numerical relationships among points in an n-dimensional space.
  • Word2vec: Word2vec made smaller, information-rich vectors practical for NLP and helped improve machine translation and sentiment classification.Typical embeddings were approximately 100–500 dimensions and were amenable to machine-learning classifiers.
  • Transformer language models: Transformer language models extend distributional learning by predicting words from context, enabling one pretrained model to be fine-tuned across many NLP tasks.They became the underlying architecture for systems such as ChatGPT and performed better across multiple benchmarks and tasks.
  • Limitation: Text-only distributional models represent textual patterns successfully but do not possess concrete word meaning grounded in physical experience.Vision-language models add images, but images cover only a small portion of the physical experience through which humans connect meaning and language.

7 Human First Language Acquisition

The paper surveys child language learning to identify aspects of language and semantics that may matter for computational semantics. Children begin speaking early despite limited exposure, and language development is intertwined with cognition, perception, interaction, and relationships.

  • Section purpose: The survey compares how children learn language with how computational models learn language to identify potentially important aspects of computational semantics.The section draws on child development and related interdisciplinary perspectives.
  • Early development: Children begin speaking their first words at about 12 months despite exposure to only a few million words.Interactive, spoken dialogue is described as the most fundamental and natural setting for human communication.
  • Learning conditions: Child language learning requires situated interaction, shared attention, speech, and agreement about how words refer.These conditions describe the social and perceptual setting in which children acquire language.
  • Cognitive development: Piaget’s developmental stages describe changing capacities for reasoning, language, concepts, and mental simulation from infancy through adolescence.The stages run from sensorimotor development through pre-operational, concrete operational, and formal operational abilities.
  • Cognitive development: Child development matters for language because children’s changing ways of perceiving and operating on the world alter comprehension and production.Language also helps children interact with others, organize their understanding of the world, and foster relationships.

7.2 The Setting: Situated, Spoken Dialogue

The basic setting for first language learning is situated, spoken interaction in which people share attention and update language through joint activities. Attention filters perception and helps caregivers connect words with the objects children are focused on.

  • Situated interaction: Children first learn language in co-located interaction where speakers directly perceive one another in shared time and space.Although texting and email may be more common in some societies, spoken interaction is presented as the basic setting.
  • Shared attention: Shared attention lets people use pointing and visual saliency to direct one another toward relevant objects or events.The account treats extralinguistic information as part of communicative interaction.
  • Spoken dialogue: Speech is the primary medium children use before literacy, while signed language is also a valid primary medium.Many languages lack writing systems, and children learn to speak or sign before reading and writing.
  • Joint activities: Joint activities let people use language, hear language, and update their understanding through shared experience.Language learning is therefore embedded in activities with other people rather than isolated symbol processing.
  • Attention: Attention is critical because people filter out information outside their current focus, as illustrated by participants missing a gorilla while counting basketball passes.The paper notes that some details of the study were contested, but not the importance of attention.
  • Attention: Caregivers often name an object a child is attending to, and later use learned words such as spoon to redirect the child’s attention.The child’s response can include locating, picking up, and handing the object to the caregiver.

7.4 Referring to Objects

Human language develops through social interaction, incremental communication, and continual updating of shared word meanings. Conversational grounding and clarification help people repair misunderstandings and build common reference.

  • Children begin communicating by indicating objects, then rapidly learn words for objects and actions and combine them into increasingly complex phrases.
  • Caregivers support language development by repeating children’s utterances, taking dialogue turns, rephrasing them grammatically, and maintaining slightly advanced complexity.
  • Human speech is produced and understood incrementally because attention is limited and spoken information travels over time, unlike large-scale parallel processing.
  • Conversational grounding updates interlocutors’ understanding through interaction, including clarification requests when speech is unheard, ambiguous, or misunderstood.
  • 3.5% of dialogue turns in one corpus involved communication breakdowns that resulted in clarification requests, while spontaneous dialogue showed 3–6%.
  • Conversational grounding can scaffold symbol grounding when interaction links a word to an object and its visual properties.

7.5 Affordances

Human object meaning includes both perceptual properties and affordances—the actions objects make possible. Language models can acquire affordance knowledge from text, but textual description may not establish human-like meaningfulness.

  • Humans understand objects through both surface properties, such as shape and color, and their possible uses.
  • An affordance describes what an object can do or enable, such as chairs enabling sitting and ladders enabling climbing.
  • Linguistic meaning is tied to the actions objects make possible rather than being solely an intrinsic property of the objects.
  • Language models can acquire knowledge of object affordances when those affordances are described in their training text.
  • Whether text-derived affordance knowledge makes objects meaningful to language models remains an open question in the discussion.
  • Curiosity and exploration are presented as precursors to language learning because children investigate the world they must connect language to.

7.7 Intention

Intention connects meaning to purposeful action: agents pursue possible outcomes, monitor progress, and distinguish actual from possible events. Understanding intentions therefore supports language use and interpretation.

  • Intent is defined as choice with commitment and is associated with acting to achieve outcomes in the world.
  • Intentions normally pose problems, constrain adoption of other intentions, require success tracking, and distinguish possible from actual events.
  • If an agent intends a possible outcome, its beliefs concern the outcome’s possibility, current non-achievement, and conditions under which it can achieve it.
  • Agents need not intend every expected side effect of their intentional actions.
  • A hunger example illustrates changing behavior from sitting inactive to obtaining food in order to bring about an intended outcome.
  • Children can recognize another person’s goal and help achieve it nonverbally, as shown by toddlers opening a cabinet for someone carrying books.
  • Because communication and socialization involve intentions, understanding what people say requires understanding their intentions to some degree.

7.8 Theory of Mind14

Theory of Mind is the capacity to recognize, represent, and infer other people’s desires, beliefs, and intentions. It is treated as an important component of human cognition for meaningful language interaction.

  • Human Theory of Mind involves recognizing, representing, and inferring other people’s desires, beliefs, and intentions.
  • Theory of Mind has been studied in child development, cognitive, and psychological research and applied to interaction between people and machines.
  • Researchers have modeled Theory of Mind, and some have argued that language models possess a degree of it.
  • Theory of Mind helps humans understand one another and is described as necessary for learning and generating meaningful language.

8 Cognition, Emotion, and Embodiment

Human language develops through intertwined emotional, cognitive, and embodied experience rather than through cognition alone. Text-only models can approximate emotional language but do not capture its embodied grounding.

  • Cognition and Emotion: Infants communicate emotionally before acquiring meaningful spoken words, progressing from cries and smiles to intentional vocal sounds and first words.
  • Cognition and Emotion: Many word meanings include positive or negative emotional valence, including abstract terms such as democracy and career.
  • Cognition and Emotion: Emotional and cognitive development are inseparable, with each supporting the other during language learning and broader development.
  • Embodiment: Text-based models may infer emotional content, but their representations do not capture the embodied emotional content that forms part of word meanings.
  • Embodiment: Human linguistic meaning is grounded in multimodal bodily experience, sensory inputs, action, cognition, and emotion.

9 Conclusion: Understanding Meaning

The paper argues that language models process symbolic language and can understand it abstractly, but their lack of embodied experience limits deeper meaning. It proposes integrating formal, grounded, and distributional semantics as a possible path forward.

  • Words are meaningful through their relation to the world and through personal experience, including perception, action, theory of mind, and emotional valence.
  • Language models can process symbolic communication and discuss experiences such as thirst relief, but they have not experienced those states themselves.
  • Text-only language models can understand language abstractly, yet lack of embodied and emotional experience disadvantages their understanding of deeper word meanings.
  • The paper identifies the symbol grounding and neuro-symbolic problems as unresolved challenges for computational semantics.
  • A proposed path forward combines symbolic representations grounded in the physical world with distributional representations and multimodal attention.
Loading 2608.25022v1…