140 likes | 442 Views
What is word sense disambiguation good for?. Adam Kilgarriff. The goal of WSD research is usually taken to be disambiguation between senses given in a dictionary, thesaurus or similar. senses. Word senses are a response to constraints imposed by: tradition the printed page compactness
E N D
What is word sense disambiguation good for? Adam Kilgarriff
The goal of WSD research is usually taken to be disambiguation between senses given in a dictionary, thesaurus or similar. senses
Word senses are a response to constraints imposed by: • tradition • the printed page • compactness • a single, simple method of access • resolving disputes about what a word does and does not mean
A dictionary must draw a line around the meaning.
For reference: I don't believe in word senses (Kilgarriff, 1997)
WSD application domains: • Information Retreival • Machine Translation • Parsing • Lexicography • Language Understanding
IR Work Conclusion Compared Results Krovetz and Croft (1992) a perfect WSD program would improve performance by 2%. WS-ambiguity causes only limited degradation of IR performance IR performance „with ambiguity“ and „without ambiguity“ introducing extra ambiguity did little to degrade performance the performance of [IR] systems is insensitive to ambiguity but very sensitive to erroneous disambiguation pseudoword (banana-kalashnikov) pretending to be a single word with 2 meanings Sanderson (1994) system performance with the disambiguation module improved by up to 4.3% 4% is a significant improvement Schütze (1997) sensediscrimination
MT • Two variants of ambiguity: • monolingual ambiguity • translational ambiguity No recent WSD work is employed in MT systems
Parsing Consider a case of syntactic ambiguity (PP attachment): 1 I love baking cakes with friends. 2 I love baking cakes with butter icing. Lexical information can resolve many syntactic ambiguities without being sense-disambiguated.
This good-for-nothing WSD Does WS ambiguity cause problems for NLP applications? IR: yes, to some moderate degree. Problems can substantially be overcome by using longer queries. MT: yes. Huge problem. Addressed to date by lots and lots of selection restrictions. Parsing: not known. Lexicography: yes, WSD would be of benefit. NLU: not much. NLU applications are mostly domain specic, and have some sort of domain model.