Word sense disambiguation – short Q&A
20 questions and answers on word sense disambiguation, covering sense inventories, Lesk algorithm, knowledge-based and supervised WSD and evaluation.
What is word sense disambiguation (WSD)?
Answer: WSD is the task of determining which sense of a polysemous word is intended in a given context, for example choosing between “bank” as a financial institution or the side of a river.
What is a sense inventory in WSD?
Answer: A sense inventory is a catalog of possible meanings for words, such as WordNet synsets or dictionary definitions, against which WSD systems choose the most appropriate sense for each occurrence.
How does the Lesk algorithm perform WSD?
Answer: The Lesk algorithm selects the sense whose dictionary definition has the largest word overlap with the context of the target word, assuming more shared words indicate a better semantic match.
What is knowledge-based WSD?
Answer: Knowledge-based methods use lexical resources like WordNet or semantic networks to exploit glosses, relations and graph structure for disambiguation without requiring sense-labeled training data.
What is supervised WSD?
Answer: Supervised WSD treats each sense as a label and trains classifiers on sense-annotated corpora, using contextual features or embeddings of surrounding words to predict the correct sense.
Why is supervised WSD challenging in practice?
Answer: High-quality sense-annotated data is expensive and limited, sense inventories are fine-grained, and sense distributions are skewed, making it hard to train robust models for all words and senses.
What is the most-frequent-sense baseline in WSD?
Answer: The most-frequent-sense baseline always predicts the sense that occurs most often for each word in a training corpus; surprisingly, it is strong and difficult to beat consistently across tasks.
How is WSD evaluated?
Answer: WSD is typically evaluated using accuracy over sense-annotated test sets, comparing predicted senses against gold labels, sometimes alongside baselines like most-frequent-sense or Lesk variants.
What role does WordNet play in many WSD systems?
Answer: WordNet provides synsets, glosses, semantic relations and example sentences that underpin knowledge-based algorithms, serve as sense inventories and support feature extraction for supervised WSD models.
How do neural contextual embeddings impact WSD?
Answer: Contextual embeddings from models like BERT yield different vectors for different occurrences of the same word, naturally encoding sense distinctions and enabling simple classifiers to achieve strong WSD performance.
What is sense representation learning?
Answer: Sense representation learning aims to derive separate embeddings for each sense of a word, either by clustering contextual usages or by aligning contexts with sense inventories, improving WSD and lexical semantics tasks.
What is the difference between homonymy and polysemy in WSD?
Answer: Homonyms are unrelated meanings that share form (e.g. “bank” as institution vs. river bank), while polysemous senses are related variations; WSD systems usually treat both as separate senses to be disambiguated.
Why is WSD considered a hard problem?
Answer: Human sense distinctions can be subtle and context-dependent, sense inventories may not align with intuitive meanings and annotated data is limited, making it difficult to learn generalizable sense decision boundaries.
How does WSD benefit downstream NLP applications?
Answer: Correctly resolving senses improves tasks like machine translation, information retrieval, question answering and semantic search, where misinterpreting word meaning can change the intended output significantly.
What is unsupervised WSD?
Answer: Unsupervised WSD clusters word occurrences based on contextual similarity without using sense-labeled data, then attempts to align clusters with dictionary senses or use them directly as induced sense categories.
What are all-words vs. lexical-sample WSD tasks?
Answer: In all-words WSD, systems must disambiguate every content word in a text, while lexical-sample tasks focus on a predefined subset of target words with many labeled examples for each.
How do graph-based WSD methods work?
Answer: Graph-based methods build graphs where nodes are senses or words and edges reflect semantic relatedness; algorithms like PageRank or random walks identify senses that are most central or best connected to the context.
What is sense inventory granularity and why does it matter?
Answer: Granularity refers to how finely senses are split; very fine-grained inventories are harder for both humans and systems to disambiguate reliably, affecting annotation agreement and model performance.
How do multilingual and cross-lingual WSD approaches work?
Answer: They exploit parallel corpora, multilingual embeddings or shared sense inventories across languages, using translations or aligned contexts to help infer senses where monolingual resources are limited.
Are large language models solving WSD implicitly?
Answer: Large language models capture many sense distinctions through context, often achieving strong WSD accuracy without explicit sense labels, but explicit WSD remains useful for interpretability and knowledge integration.
🔍 WSD concepts covered
This page covers word sense disambiguation: sense inventories, Lesk algorithm, knowledge-based and supervised WSD, contextual embeddings, evaluation and applications in downstream NLP tasks.