The mental lexicon is the human cognitive system that makes words available for recognition, interpretation, and production. It links knowledge of a word's spoken and written forms with its meanings, grammatical properties, and relationships to other words. When a listener recognizes a voice saying "garden," a reader understands the letters on a page, or a speaker retrieves "garden" while describing a scene, lexical processing coordinates different kinds of information with remarkable speed. The scientific question is how that information is represented, learned, activated, selected, and retrieved. The mental lexicon is best understood as a dynamically organized set of linguistic representations and processing mechanisms rather than a literal alphabetical dictionary stored in the brain. The strongest evidence comes from reaction-time studies, priming experiments, speech errors, learning studies, neurophysiology, and computational modeling. Levelt, Roelofs, and Meyer (1999); Stella et al. (2024).
What Is the Mental Lexicon?
A mental lexicon is a theoretical construct in psycholinguistics: the organized lexical knowledge and procedures that enable a person to connect linguistic forms with meaning and use them in context. A lexical entry is a useful modeling term for a representation associated with a word or word-like expression. Whether the brain literally implements discrete entries, distributed representations, or a mixture remains an active theoretical question. Linguistic knowledge is structured: a person can know how to pronounce "run," recognize its spelling, distinguish its senses, combine it with grammatical endings, and relate it to "jog," "race," and "operate" without consulting a conscious list.
The distinction between lexical knowledge and lexical access is central. Lexical knowledge concerns representations acquired through experience. Lexical access concerns the operations by which an acoustic pattern, written string, communicative intention, or contextual cue activates information associated with a word. Lexical retrieval usually emphasizes selecting and recovering information needed to produce a word, although the expression is also used more broadly in comprehension. Word recognition is often the outcome of successful access from perceptual input. These processes interact over time, but experiments can separate their measurable contributions.
A mental lexicon is organized by several relationships
Words can be related by meaning, sound, spelling, morphology, syntactic category, familiarity, and personal experience. "Dog" and "cat" share semantic features and occur in related contexts. "Dog" and "fog" share phonological and orthographic structure. "Teach" and "teacher" share a morphological family. "Bank" links one familiar form to more than one conventional sense. A lexical model must explain how these partially overlapping relationships shape successful access and occasional interference. Network models encode some of these relationships as nodes and links; distributed models represent a word through a pattern of activity over many units. These are scientifically useful descriptions at different levels of abstraction. Steyvers and Tenenbaum (2005); Stella et al. (2024).
What counts as a word?
Vocabulary size depends on the counting unit. A word form is a particular realization, such as "walk," "walks," or "walked." A lemma groups inflected forms around a lexical base and its grammatical behavior. A word family can additionally contain derivations. A multiword expression, such as "by and large," may have properties that exceed the simple sum of its words. Proper names, technical terms, inflections, and partially known words further complicate measurement. One large study estimated approximately 42,000 lemmas for an average 20-year-old American English speaker under its specific operational definitions; that result is an empirical estimate tied to its sample and scoring method, not a fixed human capacity. Brysbaert, Stevens, Mandera, and Keuleers (2016).
What Information Is Represented About a Word?
Meaning and conceptual relationships
A lexical representation supports access to one or more meanings, but lexical semantics and general conceptual knowledge overlap in complicated ways. Understanding "apple" draws on linguistic convention and knowledge of objects, colors, actions, food, and social situations. In some models, word forms link to conceptual representations stored in a broader semantic system. In other models, meanings are encoded as patterns distributed across interacting features. The distinction is analytically useful because knowing a concept and retrieving the corresponding name can dissociate. An individual may recognize an object and demonstrate its use while temporarily failing to say its name. That observation motivates models with partially separable form and meaning processes; it does not locate every meaning in a separate brain compartment. Levelt et al. (1999); Stille et al. (2020).
Phonological forms and spoken words
Knowing a spoken word includes information about speech sounds, their order, syllabic structure, prosodic possibilities, and how a word varies across voices and speaking styles. Speech arrives over time, usually without clear pauses marking every word boundary. A listener therefore activates plausible lexical candidates before an utterance is complete. Hearing the beginning of "candle" can temporarily support other words with similar beginnings; later information reduces the candidate set. This motivates the distinction between generating candidates and resolving competition. Marslen-Wilson (1987); McClelland and Elman (1986).
Orthography and written word forms
Literate readers acquire systematic relationships among letters, letter sequences, pronunciation, and meaning. Written lexical processing is influenced by spelling regularities, visual similarity, and word frequency. Orthographic neighborhood refers to the set of written forms similar to a target by a specified criterion; its effect depends on the task and the reader. Reading a familiar word can involve parallel activation of orthographic, phonological, and semantic information, with relative contributions varying by writing system and learning history. A mental lexicon therefore develops differently when literacy is acquired, although spoken lexical knowledge exists before conventional reading. Seidenberg and McClelland (1989); Yap, Balota, Sibley, and Ratcliff (2012).
Grammar, morphology, and combinatorial constraints
A word also carries grammatical affordances. Verbs differ in the arguments they tend to take; nouns vary in countability; affixes signal tense, number, derivation, or grammatical class. The term lemma often refers to lexical information sufficient for grammatical planning before the phonological word form has been fully encoded. Its exact theoretical implementation differs by production model. Morphological structure matters because a familiar complex word may be accessed through its whole form, through constituents such as stems and affixes, or through processes combining both routes. Levelt et al. (1999); Taft and Forster (1975).
How Are Words Stored in Long-Term Memory?
Storage is a shorthand for changes in durable representations produced through learning and repeated use. It does not imply that each word occupies a single physical address. The enduring representations must preserve enough regularity to support recognition across noisy input, different speakers, grammatical contexts, and long gaps between encounters. They must also remain adaptable when a person learns a new sense, spelling, pronunciation, or usage pattern. Long-term lexical knowledge is commonly modeled through weighted connections or structures whose accessibility changes with experience.
Discrete entries, feature representations, and distributed patterns
Classical localist accounts assign functional units to words or lexical candidates and allow activation to spread among them. Connectionist models instead emphasize distributed patterns across sublexical and semantic units. Both can capture regularities in behavior, but they differ in what is explicitly stored, how similarity is computed, and how new words generalize to unfamiliar forms. A model with a single node for every word is an intelligible computational hypothesis, while evidence for lexical processing does not establish that physical neurons correspond one-to-one with entries. Seidenberg and McClelland (1989); Steyvers and Tenenbaum (2005).
Lexical networks and small-world organization
Researchers can construct semantic networks from free associations, dictionaries, word similarities, or large corpora. Such networks often exhibit clustered neighborhoods and relatively short paths connecting many words. These properties can help explain why related items facilitate one another, why some items are highly connected, and why lexical search sometimes reaches remote associations. Yet a network derived from an association survey is a measurement of one relationship under one task, not a complete anatomical map of words. Different edges represent different phenomena: association is distinct from shared semantic features, and semantic similarity is distinct from phonological overlap. Multilayer networks seek to retain these distinctions rather than compressing them into one undifferentiated connection. Steyvers and Tenenbaum (2005); Stella et al. (2024).
Learning new lexical representations
New words become accessible through repeated encounters linking sound or spelling with use and meaning. Early learning can yield familiarity without immediately producing competition against entrenched words. In experiments with novel spoken forms, researchers have found evidence that lexical integration unfolds over time and can be reflected in competition with similar known words. Such results support accounts that distinguish initial encoding from subsequent integration and consolidation, while the contribution of sleep, repetition, and context depends on design and material. Gaskell and Dumay (2003).
Lexical learning is cumulative and selective. A child may know how a word sounds before commanding its full grammatical or semantic range; an adult may recognize a technical term in print but hesitate to pronounce it. Repeated use deepens associations among form, meaning, and contexts. This is why receptive knowledge, productive knowledge, and speed of retrieval are related but empirically distinguishable outcomes. A vocabulary test cannot automatically tell us how quickly a speaker can access a particular word during spontaneous conversation.
Lexical Access: From Input to Recognition
Activation begins before recognition is complete
When a spoken word unfolds, multiple candidates may initially fit its beginning. Acoustic evidence continuously changes the strength of their support. A word recognition model must specify how candidates are generated, how competition is resolved, what counts as sufficient evidence, and how context interacts with incoming sound. Marslen-Wilson's cohort framework describes the development and testing of candidates during spoken input. The TRACE model provides a computational interactive-activation account in which multiple levels exchange activation through excitatory and inhibitory connections. Both made explicit predictions and advanced experimental tests, while neither should be treated as the final literal anatomy of listening. Marslen-Wilson (1987); McClelland and Elman (1986).
Competition is an ordinary feature of recognition
Similar words can become partly active at the same time. Competition can arise from overlapping initial sounds, spelling patterns, meanings, or task-relevant response alternatives. An efficient lexicon therefore cannot simply retrieve the single best match without regard for other plausible candidates. Recognition emerges through accumulation, discrimination, contextual integration, and decisions whose balance depends on the experimental task. Listeners may resolve ambiguity rapidly during natural speech but display different response patterns in a lexical decision task requiring an explicit word/nonword judgment.
Semantic context and predictive expectations
Sentence context changes which meanings and word continuations are plausible. In the expression "She deposited money at the bank," the financial sense becomes strongly supported. That may influence interpretation before conscious reflection. Context can narrow competition, bias access, and help repair ambiguous or degraded input. However, a measured contextual advantage does not by itself show that the brain generated an explicit prediction of every upcoming word. Facilitation can also arise through integration after partial input, associative activation, or adjustments in decision criteria. The mechanism must be inferred from timing, task, and control conditions. Marslen-Wilson (1987); Hutchison et al. (2013).
How Are Words Retrieved for Speaking?
Producing a word begins with an intention or message and ends with an articulated form. A well-studied theoretical sequence comprises conceptual preparation, lexical selection, morphological and phonological encoding, phonetic preparation, and articulation. This sequence is a functional analysis rather than a requirement that the brain complete every operation without temporal overlap. Lexical retrieval in speaking is particularly illuminating because a person can select an appropriate meaning while the word's sound remains difficult to recover. Levelt et al. (1999).
From concepts to lexical selection
Suppose a speaker wishes to name an instrument in a picture. The object and communicative goal activate a conceptual target. Related lexical alternatives can receive activation because they share features or associations. Selection must favor a word that fits the intended message and conversational context. Depending on the model, selection may involve competition among candidates, absolute activation thresholds, or context-sensitive criteria. Semantic substitutions and category-related naming errors reveal that lexical selection is constrained by conceptual relationships, although error patterns alone cannot uniquely identify one model.
Lemma selection and word-form encoding
In influential staged accounts, a lemma represents lexical-syntactic information selected before the sound pattern is fully prepared. The subsequent form-encoding processes retrieve phonological material, assemble syllabic plans, and coordinate articulation. Picture-word interference experiments have provided evidence for separable semantic and phonological influences at different temporal offsets. Schriefers, Meyer, and Levelt reported early semantic interference and later phonological facilitation under their experimental conditions. These effects constrain theories of access but can be accommodated by different model architectures when timing and activation parameters vary. Schriefers, Meyer, and Levelt (1990); Levelt et al. (1999).
Cascading activation and feedback
Staged feed-forward models and interactive spreading-activation models offer competing explanations of how far activation proceeds before a single lexical choice is finalized. Dell's influential account models retrieval through spreading activation among linguistic levels and uses speech errors as an important test. The contrast concerns architecture and timing: whether unselected alternatives can activate form information, whether processing is cascaded, and whether phonological levels feed activation back to lexical levels. Empirical results impose constraints on these features, while the extent of feedback varies across tasks and theories. Dell (1986); Levelt et al. (1999).
Why retrieval sometimes fails
A speaker can know a word while experiencing temporary difficulty saying it. Competition, weak access to a phonological form, fatigue, task demands, and the availability of cues can affect that moment. A tip-of-the-tongue state illustrates this partial availability but belongs to a distinct experimental and memory-related search intent. The present discussion concerns the broader architecture that makes such states possible. For a dedicated account, see Tip-of-the-Tongue Phenomenon: Why Words Feel Just Out of Reach. The presence of an occasional retrieval failure is common in ordinary language use and alone does not establish a disorder.
Frequency, Familiarity, and the Speed of Access
The word-frequency effect
Frequently encountered words tend to be recognized and named faster than less common words in many controlled experiments. Frequency effects are among the most robust observations in lexical processing, but their interpretation depends on how experience is counted. A corpus frequency estimate is a property of a sampled language environment, not a direct count of one participant's lifetime exposures. Written, spoken, conversational, and specialized registers differ substantially. Comparing two stimulus sets without adequately controlling length, morphology, neighborhood characteristics, and familiarity can misattribute effects to frequency. Brysbaert and New (2009); Yap et al. (2012).
Contextual diversity and exposure distributions
A word encountered across many different contexts may have a stronger or more flexible representation than a word encountered repeatedly in only one setting. Adelman, Brown, and Quesada showed that contextual diversity was a powerful predictor of naming and lexical decision times in the corpora and datasets they analyzed; this finding challenges accounts that treat raw token counts as the sole learning variable. The broader conclusion is that experience has structure: variety, predictability, relevance, and semantic fit matter. The precise balance of frequency and diversity estimates changes with corpus construction, participant characteristics, and statistical modeling. Adelman, Brown, and Quesada (2006).
Age of acquisition and individual familiarity
Words acquired early often differ from later acquired words in frequency, concreteness, neighborhood structure, and familiarity. Age-of-acquisition effects are therefore methodologically difficult to isolate from correlated variables. Individual reading experience, education, professional vocabulary, bilingual history, and sensory conditions also change the strength of lexical access. Large lexical datasets demonstrate reliable individual differences in accuracy and decision speed. These variations are expected properties of learning histories rather than automatically clinical deficits. Yap et al. (2012); Brysbaert et al. (2016).
Semantic Activation and Priming
Semantic priming describes a change in processing of a target following a related prime. For example, a target such as "nurse" may be responded to differently after "doctor" than after an unrelated preceding word. These differences are informative because semantic relations can affect performance even when the task asks for a basic word judgment. However, priming should not be reduced to one universal process called spreading activation. Participants can use expectations, associative strategies, controlled semantic matching, or decision mechanisms, and results depend on prime duration, relatedness, response demands, and stimulus construction. Hutchison et al. (2013); Perea and Rosa (2002).
Association, semantic similarity, and co-occurrence
Two words can frequently occur together without sharing a category; other words share features without commonly co-occurring. "Bread" and "butter" are strongly associated, while "sparrow" and "robin" are taxonomically similar. An experiment that fails to distinguish association from similarity cannot attribute the result specifically to conceptual feature overlap. Corpus distribution, norms, and the participant's linguistic experience help operationalize different relations. Interpreting a priming effect therefore requires an explicit definition of the relationship under study.
Priming is evidence about activation, not a complete map
Shorter response latency is a behavioral consequence compatible with several underlying routes. It need not prove that one particular anatomical edge has transmitted activity. The Semantic Priming Project measured responses to a large set of targets under systematically controlled conditions and demonstrated the value of item-level and participant-level variation. Carefully estimated effects are more informative than a memorable single pair of words. Hutchison et al. (2013).
Morphology and Complex Words
Morphologically complex words raise a storage problem: does a speaker retain "replayed" as a whole lexical item, assemble it from "re-", "play," and "-ed," or exploit both kinds of information? Early experiments on prefixed words supported a decomposition-based account in which apparent stems affect lexical decision. Modern psycholinguistics often evaluates families of models that allow stored whole forms, morphemic constituents, and statistical regularities to contribute with different weights. A conclusion for English prefixed written words should not automatically be extended to every language, every inflection, or spontaneous speech. Taft and Forster (1975).
Morphological representation is also relevant to productivity. People can interpret or create previously unfamiliar combinations when constituent forms and grammatical patterns are recognizable. At the same time, irregular words, idioms, and semantically opaque derivations show why knowledge of morphemes alone cannot supply every meaning. A lexicon must preserve both generalizable structure and item-specific conventional knowledge. The relative importance of these sources is an empirical question tested through priming, acquisition, reading, speech production, and cross-linguistic comparisons.
Spoken and Written Lexical Access: Shared Knowledge, Different Inputs
Speech is continuous, variable, and time-dependent; print is spatially organized and learned through literacy. Spoken recognition often involves candidate activation from unfolding acoustic evidence, while written recognition requires processing graphic patterns and learned mappings to phonology and meaning. There is overlap in semantics and lexical experience, but modality-specific constraints strongly influence behavior. Results from visual lexical decision should not be treated automatically as results about spontaneous conversation. Conversely, ambiguity resolution in spoken language cannot be mapped one-to-one onto single printed words shown in isolation. Marslen-Wilson (1987); Seidenberg and McClelland (1989).
How Scientists Study the Mental Lexicon
Lexical decision and naming tasks
In lexical decision, participants judge whether a letter string or spoken stimulus is a word. The experiment can estimate how frequency, morphological structure, relatedness, or form similarity changes response time and accuracy. Naming tasks require an overt verbal response to text or a picture. Lexical decision involves the added problem of deciding whether to respond "word," and naming requires output planning, so neither measures a pure, isolated lexical event. Carefully matched controls and multiple converging tasks are essential. Yap et al. (2012).
Picture-word interference and semantic blocking
A picture can be accompanied by a distractor related or unrelated in sound or meaning. Comparing naming latencies under different stimulus-onset asynchronies helps researchers infer when conceptual and phonological information influences production. Blocking pictures from the same semantic category can introduce repeated competition, but effects may also reflect learning and task adaptation. Models must explain patterns across conditions, not merely the direction of one average effect. Schriefers et al. (1990).
Priming and associative norms
Priming paradigms test whether earlier linguistic information changes later processing. The apparent simplicity of a prime-target pair conceals design decisions about semantic relation, prime awareness, duration, attention, response proportions, and participant expectations. Association norms and corpus-based measures help specify stimulus relations. The most credible inferences use multiple primes and targets, participant-level analysis, item-level controls, and appropriate statistical generalization. Hutchison et al. (2013).
Large-scale megastudies and corpus measures
Megastudies collect many trials over broad sets of words, enabling researchers to compare candidate predictors rather than selecting a few handcrafted examples. The English Lexicon Project illustrates the advantages of dense item-level observations and the importance of individual differences. Corpus frequency resources such as SUBTLEX-US supply reproducible exposure proxies, although corpus composition may favor one register over another. Predictive models should be tested on held-out items or independent datasets when feasible, with attention to correlated variables and measurement error. Yap et al. (2012); Brysbaert and New (2009).
Brain imaging and electrophysiology
Electroencephalography and magnetoencephalography offer temporal evidence about lexical and phonological processing; functional MRI identifies larger-scale networks associated with tasks but does not directly read out stored dictionary entries. Meta-analytic work on word production has linked cognitive components to plausible temporal and spatial patterns. The evidence supports a distributed neural organization involving interacting systems for representations, comprehension, and output control. Specific regions and timing depend on task and methodology; a diagram of localized stages is a hypothesis-generating abstraction. Indefrey and Levelt (2004); Hickok (2012).
Computational models as falsifiable explanations
A computational model specifies units, representations, parameters, and operations sufficiently clearly to generate predictions. Cohort approaches, interactive activation, staged production models, connectionist networks, and multilayer semantic graphs emphasize different mechanisms. Good evaluation asks which observations the model predicts, whether parameters generalize across datasets, and which rival explanations remain viable. A model fitting reaction times on one stimulus set need not establish the cognitive reality of every unit it contains. McClelland and Elman (1986); Levelt et al. (1999); Stella et al. (2024).
Main Theoretical Models: What Each Explains
Cohort and incremental recognition models
Cohort theories emphasize lexical candidates consistent with the beginning of a spoken sequence and describe the processes of access, selection, and integration. Their major contribution is to treat word recognition as incremental and competitive. Assumptions concerning the role of context and how mismatching candidates are eliminated have evolved. Different versions should be evaluated on their own predictions rather than treated as a single timeless doctrine. Marslen-Wilson (1987).
Interactive activation and TRACE
TRACE implements multiple levels of activation and inhibition to explain how speech input can interact with candidate word representations. Its simulations make precise claims about continuous processing and interactions. The model's success at capturing particular phenomena does not entail that every hypothesized feedback connection is biologically established. It is particularly valuable as a contrast with strictly bottom-up accounts and as a framework for asking what different levels contribute to perception. McClelland and Elman (1986).
Staged lexical production and WEAVER++
Levelt, Roelofs, and Meyer developed a theory and computational implementation that distinguishes lexical selection from later word-form planning. It organizes evidence from naming latencies, interference, and speech production. Its analytical clarity has been influential, but the extent of cascading activation and feedback is disputed, and observed time windows are approximate descriptions of tasks rather than universal biological timestamps. Levelt et al. (1999).
Spreading activation and connectionist production
Spreading-activation accounts, including Dell's, represent interactions among units and use speech-error patterns to test assumptions about the architecture of production. Distributed learning models explain word recognition through adjustment of weighted connections rather than a fixed lookup table. These approaches illuminate similarity, generalization, errors, and graded activation. Important remaining questions concern the representation of syntax and morphology, learning history, task switching, and generalization beyond laboratory vocabularies. Dell (1986); Seidenberg and McClelland (1989).
Multilayer lexical networks
Recent multilayer network proposals model distinct relations among the same words across phonological, semantic, syntactic, and other layers. The approach can represent how different forms of neighborhood structure jointly influence learning and retrieval. This is a promising modeling framework, not a verified comprehensive description of all human lexical processes. Theoretical utility should be assessed through preregistered or independent tests where possible and by comparison with alternative models. Stella et al. (2024).
Mental Lexicon, Semantic Memory, and the Brain
Lexical information is closely associated with semantic knowledge. Yet a semantic system can support concepts beyond verbal labels, and language can encode grammatical and phonological information that general world knowledge alone does not supply. The distinction is most valuable when investigating how a known meaning becomes linked to a form, or why a known form sometimes fails to activate a specific meaning. General organization of facts, concepts, and knowledge belongs to the independent article Semantic Memory: Facts, Meaning, and Knowledge.
Neural evidence supports distributed rather than single-site explanations of lexical access. Different tasks recruit auditory, visual, semantic, selection, phonological, and motor systems. Neuroimaging and lesion research can constrain models, while reverse inference from a brain region to a single psychological operation is precarious. The present article owns the cognitive organization of word representations and retrieval, while detailed anatomical localization belongs to specialized neuroscience research. Hickok (2012); Indefrey and Levelt (2004).
Individual Differences and Multilingual Lexicons
The efficiency of lexical access varies with vocabulary depth, literacy, language experience, hearing, task demands, attention, age, and proficiency. Many bilingual and multilingual speakers manage more than one learned word-form system, sometimes with partly overlapping conceptual and lexical representations. Whether lexical selection is language-specific or includes cross-language competition depends on language pairs, proficiency, context, and model assumptions. It would be misleading to infer a universal cognitive advantage or disadvantage from bilingual word-access studies. The lexicon-related question here is the coexistence and accessibility of representations; broader bilingual language processing and executive-control effects require distinct reviews. Yap et al. (2012); Stella et al. (2024).
Vocabulary knowledge also differs by domain: specialized professional terms may be highly accessible at work and slower in everyday contexts. A learner can have strong written recognition but weak phonological retrieval for the same item. Experimental researchers should therefore characterize participant populations, languages, exposure histories, and tasks rather than assume that English-speaking university students represent lexical processing universally.
Clinical Boundaries: When Difficulty Finding Words Matters
Momentary word-finding difficulty is part of normal language performance. More persistent or impairing language difficulties can occur for many reasons, including neurological disease, developmental conditions, changes in hearing, medication effects, and other factors. A speech error, hesitancy, or slow laboratory response is not equivalent to aphasia, developmental language disorder, dyslexia, apraxia of speech, or a diagnosed clinical condition. These have distinct diagnostic criteria, mechanisms, assessment methods, and clinical ownership. This article explains cognitive lexical processes; it does not provide diagnostic screening or treatment instructions. Medical or speech-language evaluation is appropriate when a person experiences new, persistent, progressive, or functionally significant changes in speech or comprehension, particularly when onset is sudden. Causal interpretation belongs to qualified clinical assessment, not to self-diagnosis from a psycholinguistic example.
What Large Language Models Can and Cannot Show About the Human Lexicon
Computational language systems can encode statistical regularities among words and produce contextually appropriate continuations. Their success can inspire hypotheses about representation, distributional learning, and sensitivity to linguistic context. It does not establish that a model stores words as humans do, reproduces the developmental pathways of lexical acquisition, or possesses subjective linguistic experience. Comparisons are strongest when they focus on explicit behavioral benchmarks and identifiable processing hypotheses. Surface similarity between a generated sentence and human speech is insufficient to settle the cognitive mechanisms involved. Human mental-lexicon research includes real-time perception, embodied learning history, speech articulation, interactive tasks, and neurophysiological constraints that text-prediction performance alone does not reproduce. These are empirical and methodological distinctions rather than a philosophical verdict about every possible artificial system.
Established Findings, Competing Explanations, and Open Questions
The existence of durable lexical knowledge, sensitivity to frequency and experience, activation of related representations, competition among similar forms, and separable influences of meaning and phonological form are supported by multiple convergent research traditions. This evidence establishes that access is structured, graded, and context-sensitive. It does not establish a single agreed representational architecture. The precise degree of feedback between processing levels, how morphology combines stored forms and compositional computation, and how lexical networks should be interpreted neurally remain contested. The roles of contextual prediction, individual learning histories, multilingual experience, and different writing systems require designs that distinguish rival explanations.
Several issues also limit generalization. Reaction time measures decisions and motor responses in addition to lexical processing. Corpus frequency is an imperfect exposure proxy. Semantic network topology depends on the relation used to construct edges. Imaging maps correlations and task-dependent activation. Models may fit existing data without successfully forecasting new results. Large behavioral datasets improve reliability but do not eliminate theoretical underdetermination. A scientifically grounded picture therefore combines multiple methods while stating exactly what each supports. Yap et al. (2012); Stella et al. (2024).
Practical Implications for Reading, Learning, and Communication
A better understanding of lexical organization helps explain why varied meaningful encounters with a word can support learning, why repeated retrieval may improve access, and why vocabulary knowledge should be assessed through more than one task. Educational applications should distinguish acquiring a meaning, recognizing a written form, identifying a spoken form, and producing a word fluently. These are related outcomes that may develop at different rates. The evidence does not support a universal single technique for every learner, language, or age. Contextual diversity findings suggest that exposure in varied contexts deserves attention, while learning interventions require independent outcome studies before their effectiveness can be claimed. Adelman et al. (2006); Brysbaert et al. (2016).
For communication, lexical access offers an explanation of natural pauses, reformulations, semantic substitutions, and occasional word-finding failures. Vocabulary and lexical speed can differ across domains and conversational circumstances. These distinctions can encourage more accurate observation of communication while preserving the separation between ordinary variation and clinical evaluation.
Frequently Asked Questions
Is the mental lexicon the same as a dictionary in the brain?
It is a scientific model of lexical knowledge and processing, not a literal paper dictionary or proven collection of one-neuron-per-word entries. Words are connected through multiple linguistic and experiential relationships, and the way those relationships are implemented remains an active research question.
Are words stored by meaning, spelling, or sound?
All three kinds of information matter, alongside morphological and grammatical properties. Models vary in whether they posit separable entries, linked representations, or distributed patterns. Different tasks emphasize different sources of information.
Why can I recognize a word but fail to say it?
Recognition and production impose different demands. Knowing an intended meaning does not guarantee rapid recovery of its phonological form in every situation. The phenomenon is compatible with models distinguishing conceptual, lexical, and word-form processing, although a momentary failure cannot identify its own mechanism.
Does using a word more often always make it easier to retrieve?
Experience usually matters, but frequency is not the whole story. Diversity of contexts, the word's form, competing neighbors, the task, and individual history also influence performance. A frequency effect observed in one corpus or task does not translate directly into a person's exact retrieval time.
Do words have separate locations in the brain?
Neural evidence points to interacting, distributed systems involved in visual and auditory word recognition, semantics, production planning, and articulation. Models may employ localized functional units, but those units are not literal proof of single anatomical locations for individual words.
Can semantic priming prove that words are stored in a network?
Priming provides evidence that relations between words affect processing. It is compatible with multiple architectures and can include strategic and decisional influences. Network models are one family of explanations tested against that evidence.
Are bilingual speakers using two completely separate mental lexicons?
The degree of separation and interaction is a theoretical and empirical question. Many models allow language-specific forms to interact with shared or overlapping meaning representations, with patterns dependent on proficiency, context, and languages involved. No simple universal division follows from bilingualism alone.
Is a large language model a mental lexicon?
The term mental lexicon describes human linguistic knowledge and access mechanisms. Computational language models provide useful comparisons and computational hypotheses, but reproducing linguistic patterns does not establish human-equivalent lexical processing or subjective experience.
Related Articles
Cluster Language, Thought and Psycholinguistics — the broader scientific chapter on human language processing and its relationships to thought, learning, and linguistic interaction.
Tip-of-the-Tongue Phenomenon: Why Words Feel Just Out of Reach — an independent explanation of temporary lexical retrieval difficulty.
Semantic Memory: Facts, Meaning, and Knowledge — conceptual and general-knowledge organization as a related but distinct memory domain.
Word Recognition: How the Mind Identifies Spoken and Written Words — the recognition processes that connect language input with lexical knowledge.
References
Adelman, J. S., Brown, G. D. A., and Quesada, J. F. (2006). Contextual diversity, not word frequency, determines word-naming and lexical decision times. Psychological Science, 17(9), 814–823. https://doi.org/10.1111/j.1467-9280.2006.01787.x
Brysbaert, M., and New, B. (2009). Moving beyond Kučera and Francis: A critical evaluation of current word frequency norms and the introduction of a new and improved word frequency measure for American English. Behavior Research Methods, 41, 977–990. https://doi.org/10.3758/BRM.41.4.977
Brysbaert, M., Stevens, M., Mandera, P., and Keuleers, E. (2016). How many words do we know? Practical estimates of vocabulary size dependent on word definition, the degree of language input and the participant’s age. Frontiers in Psychology, 7, 1116. https://doi.org/10.3389/fpsyg.2016.01116
Dell, G. S. (1986). A spreading-activation theory of retrieval in sentence production. Psychological Review, 93(3), 283–321. https://doi.org/10.1037/0033-295X.93.3.283
Gaskell, M. G., and Dumay, N. (2003). Lexical competition and the acquisition of novel words. Cognition, 89(2), 105–132. https://doi.org/10.1016/S0010-0277(03)00070-2
Hickok, G. (2012). Computational neuroanatomy of speech production. Nature Reviews Neuroscience, 13, 135–145. https://doi.org/10.1038/nrn3158
Hutchison, K. A., Balota, D. A., Neely, J. H., Cortese, M. J., Cohen-Shikora, E. R., Tse, C.-S., Yap, M. J., Bengson, J. J., Niemeyer, D., and Buchanan, E. (2013). The Semantic Priming Project. Behavior Research Methods, 45, 1099–1114. https://doi.org/10.3758/s13428-012-0304-z
Indefrey, P., and Levelt, W. J. M. (2004). The spatial and temporal signatures of word production components. Cognition, 92(1–2), 101–144. https://doi.org/10.1016/j.cognition.2002.06.001
Levelt, W. J. M., Roelofs, A., and Meyer, A. S. (1999). A theory of lexical access in speech production. Behavioral and Brain Sciences, 22(1), 1–38. https://doi.org/10.1017/S0140525X99001776
Marslen-Wilson, W. D. (1987). Functional parallelism in spoken word-recognition. Cognition, 25(1–2), 71–102. https://doi.org/10.1016/0010-0277(87)90005-9
McClelland, J. L., and Elman, J. L. (1986). The TRACE model of speech perception. Cognitive Psychology, 18(1), 1–86. https://doi.org/10.1016/0010-0285(86)90015-0
Perea, M., and Rosa, E. (2002). The effects of associative and semantic priming in the lexical decision task. Psychological Research, 66, 180–194. https://doi.org/10.1007/s00426-002-0086-5
Schriefers, H., Meyer, A. S., and Levelt, W. J. M. (1990). Exploring the time course of lexical access in language production: Picture-word interference studies. Journal of Memory and Language, 29(1), 86–102. https://doi.org/10.1016/0749-596X(90)90011-N
Seidenberg, M. S., and McClelland, J. L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96(4), 523–568. https://doi.org/10.1037/0033-295X.96.4.523
Stella, M., Citraro, S., Rossetti, G., Marinazzo, D., Kenett, Y. N., and Vitevitch, M. S. (2024). Cognitive modelling of concepts in the mental lexicon with multilayer networks: Insights, advancements, and future challenges. Psychonomic Bulletin and Review, 31, 1981–2004. https://doi.org/10.3758/s13423-024-02473-9
Steyvers, M., and Tenenbaum, J. B. (2005). The large-scale structure of semantic networks: Statistical analyses and a model of semantic growth. Cognitive Science, 29(1), 41–78. https://doi.org/10.1207/s15516709cog2901_3
Stille, C. M., et al. (2020). Modeling the mental lexicon as part of long-term and working memory and simulating lexical access in a naming task including semantic and phonological cues. Frontiers in Psychology. https://pubmed.ncbi.nlm.nih.gov/32774315/
Taft, M., and Forster, K. I. (1975). Lexical storage and retrieval of prefixed words. Journal of Verbal Learning and Verbal Behavior, 14(6), 638–647. https://doi.org/10.1016/S0022-5371(75)80051-X
Yap, M. J., Balota, D. A., Sibley, D. E., and Ratcliff, R. (2012). Individual differences in visual word recognition: Insights from the English Lexicon Project. Journal of Experimental Psychology: Human Perception and Performance, 38(1), 53–79. https://doi.org/10.1037/a0024177