WORD SENSE INDUCTION USING DOMAIN-SPECIFIC CORPORA
James A. Boyden, James Curran · 2010
When we talk about a mouse, do we mean a rodent or a computer device? When we talk about a star, do we mean an actor or a point of light in the sky? Words can have multiple senses; without knowing which sense of a word was intended, we can’t interpret the overall meaning. Word sense induction is the task of discovering the senses of a word that occur in a given corpus of text. NEED FOR WORD SENSE INDUCTION • Manually-constructed word sense inventories exist, but they have problems: 1. missing senses 2. out-of-date senses 3. senses that are too fine-grained for practical usage (see Figure 1 for an example) 1. a celestial body of hot gases that radiates energy derived from thermonuclear reactions... 2. someone who is dazzlingly skilled in any field 3. any celestial body visible (as a point of light) from the Earth at night 4. a plane figure with 5 or more points; often used as an emblem 5. an actor who plays a principal role 6. a performer who receives prominent billing 7. a star-shaped character * used in printing 8. the topology of a network whose components are connected to a hub Figure 1: The 8 senses of star given in WordNet 2.0 DISTRIBUTIONAL SIMILARITY • Distributional similarity can calculate the similarity of the meanings of words, based upon the contexts in which the words appear [Lin, 1998]: I fed the dog mouse pigeon snake