Lexical Substitution — Context-Sensitive Word Replacement
Lexical Substitution (Context-Sensitive Word Replacement) · Also known as: sözcüksel ikame, Sözcüksel İkame (Lexical Substitution), context-aware synonym replacement, word-level paraphrasing
Lexical substitution is a natural-language-processing task — formalised by McCarthy and Navigli through the SemEval shared task series starting in 2007 — that replaces a target word in a sentence with a semantically equivalent alternative that preserves the meaning of the surrounding context. It draws on synonym resources such as WordNet or on distributional word embeddings and masked language models to generate and rank candidate replacements, and is used for text robustness testing, style adaptation, and training-data augmentation.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Lexical substitution applies whenever you have a text corpus and either a WordNet-compatible lexical resource or a distributional word embedding model available for the corpus language. It is appropriate for generating paraphrases for data augmentation, for producing stylistically adapted variants of documents, and for stress-testing NLP pipelines against surface-level lexical variation. Because the method requires a synonym resource or a language model, it cannot run on a language for which neither resource exists. The minimum corpus size is ten documents; there is no normality assumption because the method operates on text, not numeric measurements.
Strengths & limitations
- Preserves the grammatical structure and core meaning of the original sentence, producing natural-sounding output.
- Works with multiple backend resources (WordNet, word embeddings, or BERT fill-mask), making it adaptable to the available infrastructure.
- Effective for low-cost training-data augmentation: many diverse paraphrases can be generated without manual rewriting.
- No normality assumption — applies to any language for which a synonym resource or language model exists.
- Quality depends directly on the synonym resource or language model: if WordNet coverage is sparse for the target domain, candidates will be poor.
- Word-sense disambiguation errors upstream propagate into candidate generation, producing substitutes that fit the wrong sense.
- Context window is limited: very long-range pragmatic constraints (e.g., narrative consistency across paragraphs) are not captured.
- Does not handle multi-word expressions or idioms gracefully — substituting one word inside a fixed phrase typically breaks the expression.
Frequently asked
What is the difference between lexical substitution and simple synonym replacement?
Simple synonym replacement looks up a word in a thesaurus and picks any entry, ignoring context. Lexical substitution first disambiguates the word sense in the specific sentence and then ranks candidates by how well they fit that context — so the output is coherent in context, not just lexically related.
Do I need BERT to run lexical substitution?
No. BERT fill-mask is one backend; WordNet synsets filtered by distributional similarity are another. If you have a WordNet for your language, a WordNet-based pipeline works without any neural model. BERT-based pipelines tend to produce higher quality substitutes on languages with large pretrained models.
Can lexical substitution be applied to languages other than English?
Yes, provided a synonym resource (a WordNet variant, multilingual embeddings, or a multilingual masked language model) covers the target language. Applying an English-trained model or the English WordNet to another language will produce poor results.
How do I evaluate the quality of substitutions?
The standard evaluation protocol from McCarthy and Navigli (2009) compares system-ranked candidates against human-annotated gold substitutes using metrics such as recall at rank 1 and generalised average precision. For applied use, automatic fluency scoring or a sample human review of output sentences is a practical alternative.
Sources
How to cite this page
ScholarGate. (2026, June 1). Lexical Substitution (Context-Sensitive Word Replacement). ScholarGate. https://scholargate.app/en/text-mining/lexical-substitution
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- BERT EmbeddingsText mining↔ compare
- Named Entity RecognitionText mining↔ compare
- Paraphrase DetectionText mining↔ compare
- Sentiment AnalysisText mining↔ compare