Agirre and Martinez (Knowledge Sources for Word Sense Disambiguation) distinguish ten different types of information that can be useful for WSD (Word Sense Disambiguation):
- Part of speech (POS) is used to organize the word senses. For instance, in WordNet 1.6 handle has 5 senses as a verb, only one as a noun.
- Morphology, specially the relation between derived words and their roots. For instance, the noun agreement has 6 senses, its verbal root agree 7, but not all combinations hold.
- Collocations. The 9-way ambiguous noun match has only one possible sense in “football match”.
- Semantic word associations, which van be further classified as follows:
a Taxonomical organization, e.g. the association between chair and furniture. b Situation, such as the association between chair and waiter. c Topic, as between bat and baseball. d Argument-head relation, e.g. dog and bite in “the dog bite the postman”. These associations, if given as a sense-to-word relation, are strong indicators for a sense. For instance, in “The chair and the table were missing” the shared class in the taxonomy with table can be used to choose the furniture sense of chair.
- Syntactic cues. Subcategorization information is also useful, e.g. eat in the “take a meal” sense is intransitive, but it is transitive in other senses.
- Semantic roles. In “The bad new will eat him” the object of eat fills the experiencer role, and this fact can be used to better constrain the possible senses for eat.
- Selectional preferences. For instance, eat in the “take a meal” sense prefers humans as subjects. This knowledge type is similar to the argument-head relation (4d), but selectional preferences are given in terms of semantic classes, rather that plain words.
- Domain. For example, in the domain of sports, the “tennis racket” sense of racket is preferred.
- Frequency of senses. Out of the 4 senses of people the general sense accounts for 90% of the occurrences in Semcor.
- Pragmatics. In some cases, full-fledged reasoning has to come into play to disambiguate head as a nail-head in the now classical utterance “Nadia swing the hammer at the nail, and the head flew off”.
My questions are:
1) Is there other kind of information useful for WSD?
2) I only see disambiguation algorithms using at most three different kinds of information at the same time (for example A Novel Neural Sequence Model with Multiple Attentions for Word Sense Disambiguation). Is there a powerful system/algorithm integrating all those types of information?