Combining Knowledge Sources
Idea: Each partial tagger can only suggest possible senses for each word ==> It is necessary to have some method to combine the results.
Several Machine Learning Algorithm were tried:
- PROGOL by [Muggleton, 1995] --Inductive Logic Programming
- CN2 [Clark & Niblett, 1989] -- Rule Induction
- TimBL [Daelemans et al., 1998] -- Memory Based
The supervising information consists of a vector containing the senses identified by each partial tagger and 10 basic collocations. Each sense is marked as either appropriate or inappropriate and the pair of word, decision is stored. To disambiguate untagged text, the algorithm assigns the new word to the training instance closest to it and returns its class.