6533b7d6fe1ef96bd1266deb
RESEARCH PRODUCT
Building an Optimal WSD Ensemble Using Per-Word Selection of Best System
Steve LegrandHarri M. T. Saarikoskisubject
0303 health sciencesWord-sense disambiguationComputer scienceSample (material)Speech recognition02 engineering and technologyBase (topology)SemanticsSupport vector machine03 medical and health sciencesPattern recognition (psychology)Classifier (linguistics)0202 electrical engineering electronic engineering information engineering020201 artificial intelligence & image processingWord (computer architecture)030304 developmental biologydescription
In Senseval workshops for evaluating WSD systems [1,4,9], no one system or system type (classifier algorithm, type of system ensemble, extracted feature set, lexical knowledge source etc.) has been discovered that resolves all ambiguous words into their senses in a superior way. This paper presents a novel method for selecting the best system for target word based on readily available word features (number of senses, average amount of training per sense, dominant sense ratio). Applied to Senseval-3 and Senseval-2 English lexical sample state-of-art systems, a net gain of approximately 2.5 – 5.0% (respectively) in average precision per word over the best base system is achieved. The method can be applied to any base system or target word in any language.
| year | journal | country | edition | language |
|---|---|---|---|---|
| 2006-01-01 |