Here are the formal tokens of the progress of my research.



Available through CiteSeer.
Outlines for utilizing different types of lexical knowledge to the benefit of word sense disambiguation task.

Very sketchy, more of an outline of two dissertation projects.
(Co-authored with Steven Legrand.)


@ Iberamia-04

First findings using WSD system scores to predict where systems most probably go wrong. One possibly significant finding: hybrids do better with verbs than non-hybrids. It would make sense that system performance depends also on the target word and even sense.

A scratch of the surface if that proves to be true.


@ HLT-2005 (presentation)

A touch on the pre-stage of WSD: word sense discovery based on system scores. If systems frequently confuse two senses of a word, then perhaps those senses should be collated (in terms of application-specific requirements). Here machine translation grain is used as method target.

Regardless of results, using scores for both sense discovery and ensuing sense disambiguation is seen as beneficial for designing a component that contains both.


@ NLDB-05

<>Fuller description of the mySENSEVAL tool, system and database. More preliminary findings also provided. Particular emphasis on the effect of amount of training data, and how training resources for well-disambiguated words/senses could be reallocated.


@ IDDI workshop at DEXA-05

Assuming a disambiguated text, 2T is a method for automatically producing a two-term index (e.g. car-repairing, Windows-troubleshooting) out of domain text. Seen as completionary to full-text indexing, using same resources, while wreaking out another semantically meaningful 'layer' of search terms.

A precision of 90% is obtained at recall of 70%. Tested for English and Finnish, but is basically language-independent as well as light (not using syntactic trees but only lemma, POS and semantic category constraints).
        Posters New one from NLDB-05
mySENSEVAL tool operating principle and user interface
Symposium for Finnish Semantic Web ontologies project (fall 2004)
Using lexical knowledge in disambiguation. In Finnish.