Probing Lexical Ambiguity: Word Vectors Encode Number and Relatedness of Senses

Area of Science:

• Computational Linguistics
• Cognitive Science
• Natural Language Processing

Background:

• Lexical ambiguity, where words have multiple senses, impacts language acquisition and processing.
• Distributional semantics models word meaning based on context but often fail to differentiate between related and unrelated word senses.

Purpose of the Study:

• To assess if distributional meaning representations can capture the structure of lexical ambiguity, including sense relatedness.
• To evaluate distributional semantic models' ability to distinguish between monosemes, polysemes, and homonyms.

Main Methods:

• Analysis of a large sample of English words.
• Testing various distributional semantic representations for their ability to differentiate word ambiguity types.
• Comparing model performance on monosemes (unambiguous), polysemes (related senses), and homonyms (unrelated senses).

Main Results:

• Some, but not all, tested distributional semantic representations showed detectable differences between monosemes, polysemes, and homonyms.
• The study identified specific distributional models capable of capturing fine-grained aspects of word meaning structure.

Conclusions:

• Distributional semantic models can potentially capture cognitively plausible ambiguity structures.
• Future research should evaluate lexical representations not only for semantic similarity but also for their ability to model ambiguity.
• Findings contribute to understanding how computational models represent nuanced aspects of human language understanding.