Word sense disambiguation (WSD) is how a language-processing system selects the intended meaning of a word from its context. In “Sirius is the brightest star in Earth’s night,” nearby clues point to an astronomical object—not a celebrity or a shape. The system’s choice is usually made from a predefined set of meanings, so what it can distinguish depends partly on how that set was defined.
What word sense disambiguation does
Many words have more than one meaning. WSD identifies which meaning a particular use expresses, using the words around it and, when needed, broader context. A system might examine a phrase, a sentence, or a document; the useful span varies with the word and task.
As an Amazon Associate I earn from qualifying purchases.
WSD differs from word sense induction. WSD selects among senses in an existing inventory; induction seeks to discover or group senses from data rather than start with a fixed list. Nor does every modern language system contain a separate WSD component: a language model may interpret a word in context as part of its broader processing without returning an explicit sense label.
How a system chooses a sense
- Identify the target. The system locates the word whose meaning is in question.
- List candidate senses. It consults a chosen sense inventory, such as WordNet for English. WordNet groups near-synonyms into synsets, each representing a concept.
- Represent the context. The system uses surrounding words and may draw on the sentence or wider document.
- Score the candidates. Depending on the method, it compares the context with labeled examples, lexical definitions and relations, contextual representations, or a combination.
- Return an interpretation. A conventional WSD system may return a sense identifier, sometimes with a score or confidence estimate. A general language model may instead explain the intended meaning in ordinary language.
The inventory sets a boundary on the answer: a system cannot select a distinction that the inventory does not contain. Even a sound interpretation can therefore differ in specificity from the labels available to the system.
#1 Best Overall
- Used Book in Good Condition
What approaches are used, and what they trade off
Lexical-knowledge methods
Knowledge-based methods use resources such as WordNet, including definitions, semantic relations, and examples, to judge which sense fits. They can operate without a large task-specific collection of sense-labeled examples. Their reach depends on the resource’s coverage and on whether its distinctions suit the application.
Supervised learning
Supervised systems learn from text that people have labeled with senses. SemCor is a major manually sense-tagged English corpus and an important training resource. Labeled examples can teach useful contextual patterns, but coverage is a constraint: the literature notes that SemCor lacks many senses found in test sets and has few examples for some senses.
Contextual language models
Transformer models such as BERT represent a word in relation to its context and have improved results on common WSD benchmarks. Those results do not remove dependence on the benchmark’s inventory, labeling choices, or training distribution. Strong performance on familiar test data does not establish equally reliable treatment of every word or rare meaning.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Large language models
LLM-era work can ask a model to choose among senses or definitions, while a general instruction-following model may simply communicate the contextual meaning without producing a formal label. A 2026 AAAI survey reports that closed-source instruction-tuned LLMs performed comparably to specialized WSD systems in studies it reviewed. The same survey describes weaknesses with non-predominant senses and disambiguation bias in machine translation. These are findings from reviewed evaluations, not a guarantee about every model or use case.
How WSD data and benchmarks work
SemCor provides manually assigned sense labels for training. For evaluation, a widely used unified all-words benchmark brings together five datasets: Senseval-2, Senseval-3, SemEval-2007, SemEval-2013, and SemEval-2015. These datasets use WordNet senses.
A benchmark score is meaningful only in relation to its setup. Before comparing systems, check whether they share the same sense inventory, dataset, annotation granularity, and scoring method. Fine-grained distinctions can challenge annotators as well as models; training examples may also omit senses that appear in evaluation data. An aggregate score can conceal uneven performance across words and senses.
Rank #4
Why context and evaluation do not eliminate errors
- Missing or mismatched senses: A fixed inventory may not include the distinction an application needs, or its categories may fit the task poorly.
- Rare-sense coverage: A model trained mostly on common meanings may favor the predominant sense when context calls for a less frequent one.
- Limited labeled data: Some senses have few examples, and benchmark test sets can include senses not covered by the training corpus.
- Context scope: A nearby phrase may be enough in one case, while another requires the sentence or document. A system using too little context can miss useful evidence.
- Annotation choices: People may disagree about fine-grained distinctions, so a label is not always a simple, universally agreed fact.
- Domain and language shift: A result for English WordNet-based benchmarks does not by itself establish performance in another language, specialized domain, or different labeling scheme.
A 2021 survey described recent systems as surpassing an earlier 80% inter-annotator-agreement “glass ceiling” in its research context. That is a dated characterization of work in that survey, not a universal current accuracy figure or a fixed ceiling for human judgment.
How to judge a WSD result for a real task
A single leaderboard number is a weak guide if the intended use differs from the benchmark. Compare systems on the dimensions that affect the job at hand:
Best Value
- Inventory and granularity: Do the labels capture the distinctions users need?
- Training coverage: How much labeled data exists for rare or important senses?
- Context scope: Does the method use a phrase, sentence, or document?
- Evaluation setup: Were systems tested on the same data and with the same metric?
- Language and domain: Does the evaluation match the language and subject matter of the intended use?
- Output requirement: Must the system return a formal sense label, or is a useful contextual interpretation enough?
These checks also clarify why WSD remains useful in the LLM era. The 2026 AAAI survey treats it as an ongoing way to study lexical-semantic competence and model weaknesses: a fluent explanation and a consistently correct choice among fine-grained senses are not necessarily the same achievement.
Quick Recap
Sources
- Michele Bevilacqua, Tommaso Pasini, Alessandro Raganato, and Roberto Navigli, “Recent Trends in Word Sense Disambiguation: A Survey,” IJCAI 2021.
- “Analysis and Evaluation of Language Models for Word Sense Disambiguation,” Computational Linguistics 47(2), 2021.
- Roberto Navigli, “Is Word Sense Disambiguation Dead in the LLM Era?,” AAAI 2026.
- Diana McCarthy, “Word Sense Disambiguation: An Overview,” Language and Linguistics Compass, 2009.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
