Word Sense Disambiguation (WSD)

Relies on machine-readable dictionaries and lexical databases like WordNet. The classic Lesk algorithm disambiguates a word by calculating the maximum gloss overlap between its dictionary definition and the definitions of surrounding context words.

• Core Mechanism: Semantic similarity via definition matching.
• Key Resource: WordNet, BabelNet.
• Limitation: Brittle against short or noisy context; high computational cost for naive implementations.

Treats WSD as a classification problem trained on sense-annotated corpora like SemCor. Models learn to map contextual features (surrounding words, part-of-speech tags, syntactic dependencies) to the correct sense.

• Common Algorithms: Support Vector Machines (SVM), Naive Bayes, Decision Lists.
• Modern Approach: Fine-tuned transformer models (e.g., BERT) achieving state-of-the-art accuracy by modeling deep contextualized representations.
• Bottleneck: Requires expensive, manually annotated training data.

Discovers word senses automatically by clustering occurrences of a target word based on contextual similarity, without predefined sense inventories. This is known as Word Sense Induction (WSI).

• Mechanism: Context vectors are generated (e.g., via TF-IDF or neural embeddings) and grouped using algorithms like k-means or Hierarchical Dirichlet Processes.
• Advantage: Identifies domain-specific or novel senses absent from static dictionaries.
• Challenge: Mapping induced clusters to a standard sense inventory for evaluation is non-trivial.

Starts with a small set of seed examples for each sense and iteratively expands the training data from a large, unannotated corpus. The Yarowsky algorithm is the seminal example.

• Process: A classifier is trained on seeds, labels high-confidence instances, adds them to the training set, and repeats.
• Key Heuristic: 'One sense per collocation' and 'one sense per discourse'—senses are highly correlated with specific local phrases and consistent within a single document.
• Risk: Error propagation if initial seeds are noisy.

Leverages parallel corpora to resolve ambiguity by exploiting the fact that a polysemous word in one language often translates to distinct, unambiguous words in another.

• Mechanism: Aligns a source word with its translations in a bitext. If 'bank' translates to 'banque' (financial) and 'rive' (river) in French, the correct sense is determined by the target word present.
• Resource: Europarl, OPUS corpora.
• Strength: Avoids the need for manually sense-tagged data in the source language.

Disambiguates words by modeling the semantic constraints a predicate imposes on its arguments. A verb like 'eat' expects an edible entity as its object.

• Implementation: Statistical models learn the likelihood of a semantic class (e.g., FOOD) filling an argument slot, often using WordNet hypernyms or distributional similarity.
• Example: In 'He ate a bass,' the system resolves 'bass' to the fish sense because it satisfies the edible preference of 'eat,' unlike the musical instrument sense.
• Integration: Often used as a sub-module in a larger WSD pipeline.