Lesk: The Lesk Algorithm and Word Sense Disambiguation in Modern NLP

EditorialStaff Misc 30. April 2026 | 0

In the evolving landscape of natural language processing, the Lesk algorithm stands as a venerable method for word sense disambiguation, or WSD. The task of deciding which sense of a polysemous word is intended in a given context is fundamental to many NLP applications, from search and information retrieval to question answering and machine translation. The Lesk approach, often introduced under the banner of gloss overlap, offers a transparent and interpretable means of linking language to meaning by comparing contextual content with dictionary definitions. This article explores the Lesk method in depth, tracing its origins, variants, practical implementation, strengths, and limitations, and it also considers how modern NLP blends the classic Lesk idea with contemporary machine learning techniques.

What is the Lesk Algorithm?

The Lesk algorithm, named after its originator, is a method for word sense disambiguation that relies on the overlap between the context of a target word and the definitions (glosses) of candidate senses. In its simplest form, the algorithm takes a sentence or passage containing a word with multiple senses and compares the surrounding words to the definition of each sense. The sense whose gloss shares the most words or semantically related terms with the context is selected as the intended meaning. The core intuition is straightforward: the definition of the correct sense should align closely with the surrounding discourse.

Historically, the Lesk approach became a foundational baseline for WSD tasks and remains a reference point for evaluating more complex methods. It works well in scenarios where rich lexical resources, such as WordNet glosses, are available. The technique is language-agnostic in principle, provided high-quality dictionaries or glossaries exist for the target language. In practice, the Lesk algorithm is frequently implemented in its simplest form, known as the Simple Lesk, and later extended in several transformative ways to improve accuracy and robustness.

Origins of Lesk: From Concept to Computation

Origins and the early idea behind Lesk

The concept behind the Lesk algorithm emerged from the 1980s, a period that saw a surge of interest in how machines could interpret human language at a semantic level. Michael Lesk proposed a method that leveraged existing lexical resources to disambiguate word senses by measuring the overlap of word usage in context with dictionary definitions. The underlying hypothesis was intuitive: a word used in a particular sense confidently co-occurs with other terms that appear in the gloss describing that sense. In a sense, the Lesk approach makes the dictionary a predictive instrument for meaning itself.

From dictionary to context: a simple yet powerful idea

The early iterations of Lesk focused on direct overlaps between context words and the gloss terms. This simplicity contributed to both its elegance and its limitations. On one hand, the method did not require complex statistical models; on the other hand, it could struggle in short texts or when glosses were too general to capture nuances. Over time, researchers extended the approach by incorporating antonyms, examples, and hypernyms or hyponyms, creating more nuanced measures of similarity. These enhancements paved the way for a family of Lesk-inspired methods that remain influential in the field today.

How the Lesk Algorithm Works: A Step‑by‑step Guide

Understanding the mechanics of the Lesk algorithm helps illuminate why it sometimes excels while other times it falters. The method operates by comparing the lexical content surrounding a target word to the textual content of candidate senses’ glosses. The emphasis is on overlap and overlap quality. Here is a clear, practical breakdown of the classic approach and its common extensions.

Context, glosses, and the core matching principle

Suppose we encounter a sentence containing the word bank, which has several senses such as a financial institution and the side of a river. For each sense, we retrieve the gloss from a lexical resource like WordNet. We then gather the surrounding words in the sentence as the context. The Lesk algorithm computes a similarity score for each sense by counting the number of words in the gloss that also appear in the context (with potential weighting or stemming). The sense with the highest overlap is selected as the most plausible interpretation. The beauty of this approach lies in its interpretability: if the riverbank sense wins, we can trace the decision back to specific overlapping terms in the gloss and the context.

Basic (Simple) Lesk versus Extended Lesk

The Basic or Simple Lesk uses only the gloss of each candidate sense. The extended versions broaden this by incorporating additional textual material: definition examples, related senses (hypernyms/hyponyms), and even the glosses of related concepts. By enriching the information used for disambiguation, Extended Lesk can often resolve ambiguities that the simple overlap cannot. In effect, the extended versions widen the net of semantic cues the algorithm can exploit.

From overlap to similarity: variants you may encounter

Beyond mere word matches, researchers have introduced similarity measures and weighting schemes. Some variants use cosine similarity over vectorised representations of glosses and context, effectively turning Lesk into a representation learning problem. Others incorporate semantic relatedness beyond exact word matches, such as considering synonyms, derivational forms, or word embeddings to capture latent semantic connections between context and gloss terms. All these variants retain the central philosophy of using dictionary-defined senses as anchors for disambiguation, while expanding the signals used to make a decision.

Variants of Lesk: A Catalogue of Approaches

Over the years, multiple descendants of the Lesk idea have been proposed. Each seeks to address specific shortcomings, such as short context length, the sparsity of gloss content, or the need to scale to large vocabularies. Here is a survey of the most influential variants and what they bring to the table.

Simple Lesk: the classic baseline

In its simplest form, Simple Lesk computes the overlap between context words and the words contained in the gloss of each possible sense. This straightforward method provides a robust baseline that is fast to compute and easy to implement. It works particularly well when glosses are well-formed and the surrounding text offers sufficient cues. However, in short texts or highly domain-specific language, Simple Lesk can underperform compared with more sophisticated techniques.

Extended Lesk: enriching the signal

Extended Lesk broadens the source material by incorporating not only the gloss but also definitions of related senses, example sentences, and even the glosses of hypernyms and hyponyms. By leveraging a richer lexical landscape, Extended Lesk is better equipped to resolve ambiguous terms whose senses share common vocabulary, yet differ in their usage. This extension often yields more stable disambiguation results in diverse domains.

Cosine Lesk: vectorising glosses and context

Cosine Lesk introduces vector representations for glosses and context, transforming the overlap problem into a cosine similarity computation between high-dimensional vectors. Utilizing tf-idf or more modern embeddings, this approach captures partial semantic alignment even when exact word matches are sparse. Cosine Lesk demonstrates the synergy between traditional dictionary-based methods and modern vector space models, providing a bridge between classical linguistics and contemporary NLP.

Adapted Lesk: language and domain adaptations

Adapted Lesk applies domain-specific gloss modifications and language-aware scoring to improve performance in particular sectors, such as biomedicine, law, or finance. By tailoring gloss content to reflect domain terminology and typical collocations, Adapted Lesk achieves higher precision in specialised corpora. This adaptation is particularly valuable when standard lexical resources do not fully capture domain-critical senses.

Cross-lingual and multilingual Lesk

In multilingual settings, Lesk variants can exploit translations, multilingual glosses, or cross-lingual embeddings to disambiguate senses. Cross-lingual Lesk uses alignments between languages to enrich the context and gloss representations, offering improved performance in non-English texts or multilingual corpora. This variant exemplifies how the core idea of cross-referencing context with sense definitions translates across linguistic boundaries.

Graph-based and hybrid approaches

Some contemporary methods couple the Lesk philosophy with graph-based representations of lexical resources. By modelling WordNet as a graph where senses are nodes and semantic relations are edges, one can propagate evidence across related senses to sharpen disambiguation. Hybrid approaches combine the Lesk overlap with probabilistic models, neural features, or ensemble methods to achieve more robust results across varied datasets.

Using Lesk with WordNet and Other Lexical Resources

WordNet has long been the de facto lexical resource for Lesk-based WSD. Its rich network of synsets, glosses, examples, and semantic relations provides fertile ground for sense disambiguation. Beyond WordNet, other dictionaries and lexical databases—such as FrameNet, ConceptNet, or domain-specific glossaries—can be integrated to enhance coverage and accuracy. When constructing a Lesk-based system, the quality, granularity, and scope of the glosses directly influence performance. In practice, many implementations fall back to WordNet as the primary source, augmenting it with domain glosses when available.

Gloss design and the quality of definitions

The effectiveness of the Lesk approach is intimately tied to the quality of glosses. Short or overly general definitions reduce the discriminative power of the method. High-quality glosses that include meaningful content words and representative examples tend to produce stronger overlaps with contextual terms. Some projects mitigate gloss limitations by explicitly incorporating examples, phrases, or even short paraphrases into the gloss representation to increase the likelihood of meaningful overlaps.

Handling multiword expressions and lexis

WSD via Lesk can benefit from recognising multiword expressions and stable collocations that carry specific senses. By treating phrases as atomic units or by enriching glosses with common collocations, the overlap measure becomes more sensitive to actual usage patterns in text. This is especially important for words with senses that differ in part due to collocational context rather than individual word semantics.

Practical Implementation in Python: Quick Start

For practitioners, a practical implementation of the Lesk algorithm often begins with a simple Python script that leverages NLTK’s WordNet interface. The following outline sketches the essentials of a Simple Lesk implementation. It is meant as a starting point for experimentation, not a production-grade solution. You can adapt and extend it with Extended Lesk features such as hypernyms, examples, and cosine similarity with word embeddings.

import nltk
from nltk.corpus import wordnet as wn
from nltk.tokenize import word_tokenize
from nltk.corpus import stopwords
import string

def simple_lesk(context_sentence, word):
    # Tokenise and normalise
    context = [w.lower() for w in word_tokenize(context_sentence)]
    # Remove punctuation and stopwords
    context = [w for w in context if w not in string.punctuation]
    stop = set(stopwords.words('english'))
    context = [w for w in context if w not in stop]
    
    # Get senses for the target word
    senses = wn.synsets(word)
    if not senses:
        return None
    
    max_overlap = 0
    best_sense = None
    for sense in senses:
        gloss = sense.definition().lower().split()
        # Basic overlap
        overlap = len(set(gloss) & set(context))
        if overlap > max_overlap:
            max_overlap = overlap
            best_sense = sense
    return best_sense

In this snippet, the key steps are tokenisation of context, retrieval of candidate senses from WordNet, and counting the overlap between gloss words and the context. To upgrade to Extended Lesk, you would augment glosses with examples, hypernyms, and related sense information, then recalculate overlaps accordingly. If you wish to explore Cosine Lesk, you would replace the overlap counting with a vector similarity calculation, using tf-idf or embeddings to represent glosses and context as vectors.

Applications: Where Lesk Shines in the Modern NLP Landscape

Although deep learning often dominates discussions of current NLP, the Lesk algorithm still has a meaningful place in certain workflows. Here are key application areas where Lesk and its variants remain valuable tools in the toolbox of natural language understanding.

Search and information retrieval: improving query interpretation

In information retrieval, correctly identifying the intended sense of ambiguous query terms can dramatically improve relevance. A Lesk-based disambiguation step can be inserted into query processing to determine the best sense for a term before expansion or matching against indexed documents. This leads to more precise retrieval, especially for polysemous words like “bank”, “leaf”, or “record”.

Question answering and reading comprehension

Question answering systems benefit from precise sense disambiguation when parsing questions and locating relevant passages. Lesk-based methods can help align the semantic targets of questions with the correct passages, particularly when the questions involve terms with multiple senses. In some pipelines, a hybrid approach uses Lesk as a feature within a broader retrieval-and-reasoning framework.

Knowledge base augmentation and ontology alignment

When linking natural language text to structured knowledge, disambiguating word senses helps ensure correct entity recognition and relation extraction. Lesk-inspired disambiguation supports mapping surface forms to the appropriate nodes in an ontology or knowledge graph, improving data quality for downstream analytics and reasoning tasks.

Machine translation and multilingual NLP

In translation, selecting the correct sense can influence the choice of target-language word or phrase. A Lesk-based pre-processing step can correctly identify senses before translation, reducing lexical misrenders. Cross-lingual Lesk variants extend this idea by leveraging bilingual glosses to disambiguate words in multilingual corpora.

Limitations and Challenges: When Lesk Struggles

Despite its strengths, the Lesk approach has well-known limitations that teams must manage when designing NLP systems. Understanding these caveats helps practitioners decide when to apply Lesk and when to seek alternative methods or hybrids.

In texts with very short contexts, the overlap signal can be weak, making it difficult for Lesk to identify the correct sense. When a sentence is minimal, or only a few words surround the ambiguous term, gloss overlaps may be insufficient to yield a confident decision. In such cases, supplementary features or global document cues can help.

Gloss quality directly affects performance. If glosses are sparse, contain uncommon words, or fail to capture nuanced senses, the Lesk method may misfire. Conversely, high-quality glosses with rich descriptive content tend to improve outcomes. In practice, gloss resources are frequently augmented with examples or domain-specific phrases to mitigate this issue.

Language evolves, and domain-specific terminology may drift away from general-language glosses. This drift can degrade Lesk performance in specialised fields. Adaptation strategies, such as domain-adapted glosses or hybrid methods with domain-relevant knowledge bases, help address this challenge.

Words with numerous senses, such as “set” or “run,” pose a particular challenge. The large sense inventory increases the chance of overlapping gloss terms with context inadvertently, which can lead to incorrect disambiguation. Extended Lesk and vector-based enhancements can mitigate this, but careful calibration is often required.

Performance, Evaluation, and Benchmarks

Assessing WSD methods requires carefully designed evaluation. Historically, word sense disambiguation studies have used standard benchmarks such as the Senseval and SemEval datasets, which provide annotated corpora across diverse domains. The performance of Lesk variants is typically measured in terms of accuracy or F1 scores against human-annotated gold standards. In practice, Simple Lesk may yield modest yet useful improvements over baseline language models in certain tasks, while Extended and Cosine Lesk variants often outperform the basic approach in more complex contexts. It is also common to evaluate methods on domain-specific benchmarks to determine their practical viability in real-world applications.

Future of Lesk: Hybrid Methods and Deep Learning Integration

The trajectory for Lesk in the age of deep learning is not abandonment, but integration. Contemporary NLP systems frequently employ a hybrid strategy: a Lesk-like explicit disambiguation layer provides interpretable decisions anchored in lexical resources, while neural components contribute nuanced contextual understanding. For example, a model might use Cosine Lesk features as additional inputs to a neural WSD model or combine gloss-based evidence with contextual embeddings from transformers. In multilingual settings, Lesk-inspired signals can guide cross-lingual alignment and sense assignment, complementing data-driven approaches with structured linguistic knowledge.

One notable advantage of Lesk-based methods is interpretability. The decision can be traced back to explicit overlaps between context and gloss content, making it easier to diagnose errors and adjust glosses or resource coverage. In contrast, purely data-driven approaches may be more opaque, whereas a hybrid Lesk-augmented system preserves a degree of transparency that benefits error analysis and iterative refinement.

Lesk remains particularly attractive in low-resource environments where large annotated corpora for training neural models are scarce. In such contexts, leveraging lexical resources to perform disambiguation can deliver meaningful performance gains without the need for extensive supervised data. This makes the Lesk approach relevant for educational tools, smaller language communities, and specialised domains with established glossaries.

Practical Tips for Building a Lesk‑driven WSD System

Whether you are implementing Simple Lesk or an Extended Lesk variant, these practical tips can improve results and make the approach more robust in real-world pipelines.

Choose high-quality lexical resources: WordNet is a common choice for English, but domain-specific glossaries can be crucial for accuracy. Consider augmenting glosses with examples and related sense information to enrich the signal.
Preprocess context carefully: remove stopwords, apply stemming or lemmatization, and consider part-of-speech tagging to focus on content words most relevant to sense disambiguation.
Experiment with weighting: assign greater importance to content words in the context or to more informative gloss words. Simple weighting schemes can boost overlap signals without adding computational overhead.
Incorporate related senses strategically: hypernyms and hyponyms can provide helpful semantic extensions, particularly for words with broad senses.
Combine with embeddings for robustness: when overlaps are weak, vector representations of glosses and context can capture latent similarity beyond exact word matches.

A Brief Comparison: Lesk versus Other WSD Techniques

There are several families of word sense disambiguation approaches. Here is a concise comparison to place Lesk in context:

Lesk (gloss overlap) – transparent, resource-driven, interpretable, efficient for short texts but sensitive to gloss quality.
Supervised learning – high accuracy with large annotated corpora but data-hungry and less portable across domains or languages without retraining.
Unsupervised and graph-based methods – leverage structure in lexical knowledge bases, offering resilience to scarce annotations and enabling reasoning over related senses.
Neural context-aware models – excellent performance on many tasks with enough data, but often require substantial compute and can be less explainable.

Conclusion: Why Lesk Still Matters in British NLP Practice

The Lesk algorithm remains a compelling tool in the NLP toolkit precisely because it embodies a principled, interpretable approach to linking language to meaning. Its reliance on openly available glosses and its straightforward overlap logic make it accessible for researchers, educators, and practitioners alike. While modern NLP often relies on deep learning methods, Lesk-based strategies continue to offer value—especially in low-resource settings, domain-specific applications, and multilingual ventures where gloss quality and interpretability are prized.

As the field advances, the Lesk idea continues to inspire hybrid architectures that fuse explicit lexical knowledge with the nuanced generalisation of neural models. By understanding the core mechanism of the Lesk algorithm—context versus gloss overlap—you gain insight into one of language technology’s most enduring questions: how do words carry meaning within the tapestry of human discourse? In that journey, Lesk remains a reliable compass, guiding researchers toward more accurate disambiguation and clearer linguistic understanding.

Glossary of Key Terms

Below is a quick glossary to help readers navigate the terminology commonly encountered when studying the Lesk algorithm and word sense disambiguation:

: The algorithm centred on overlap between context and gloss when disambiguating word senses.
(WSD): The task of determining which sense of a polysemous word is used in a given context.
: A dictionary definition or explanation of a word’s sense, used as a semantic anchor in Lesk.
(sense): A particular meaning of a word as represented in a lexical resource such as WordNet.
: Higher-level or lower-level related senses used to enrich glosses in Extended Lesk.

Reinvigorating Lesk: A Small Example in Context

Consider a short sentence: “She deposited money at the bank after evaluating the interest rates.” Using Simple Lesk with WordNet glosses, the algorithm would compare the context words such as “deposited,” “money,” “interest,” and “rates” against the glosses for bank senses. The sense tied to a financial institution should show stronger overlap due to terms like “money” and “interest” appearing in the relevant gloss and examples. If the context instead included phrases like “towpath along the river,” the riverbank sense would likely win. Such a straightforward illustration showcases how the Lesk approach translates linguistic cues into a sense decision, reinforcing the interpretability of the method.

More Examples and Practical Scenarios

In practice, you might encounter ambiguous terms in specialised texts, such as medical literature or legal documents. Here, the precision of gloss content becomes critical. Enhanced Lesk variants—Extended Lesk or Cosine Lesk with domain-adapted glosses—often outperform the basic approach in these settings. Moreover, when you combine this with domain-specific embeddings, you gain a robust mechanism to capture both explicit term matches and latent semantic relationships. This multi-faceted strategy is particularly useful for document summarisation, question answering, and cross-domain information retrieval where precise sense assignment pays dividends.

Final Thoughts: Keeping the Lesk Spirit Alive

The Lesk algorithm embodies a spirit: that the meaning of language can be traced through definitional content and contextual cues. In a world where NLP is increasingly dominated by large neural models, revisiting and refining classic methods like Lesk provides balance. The technique remains valuable not just as a baseline, but as a practical, interpretable component that can be adapted, extended, and integrated into modern pipelines. By embracing both the simplicity of Simple Lesk and the power of Extended Lesk, Cosine Lesk, and hybrid strategies, you can build robust Word Sense Disambiguation systems that are both effective and explainable.

Lesk: The Lesk Algorithm and Word Sense Disambiguation in Modern NLP