Kevin Duh

UW, Dept of EE

Semi-supervised learning for Ranking

The problem of ranking, whose goal is to predict an ordering over a set of objects, is a key problem in many applications. In web search, for instance, ranking algorithms are used to order webpages in terms of relevance to the user. In complex NLP systems (e.g. machine translation, parsing), a set of candidate hypotheses is re-ranked such that the best one emerges at the top. While ranking has become an active research area, most work is done under the framework of supervised learning.

In this talk, I will discuss whether ranking algorithms can be extended to the semi-supervised learning framework, i.e., "Can additional unlabeled data be exploited to improve ranking performance?" I will begin by surveying possible approaches in this open area (based on prior work in semi-supervised classification). Then I will introduce a general "transductive/local" meta-algorithm for turning a standard supervised algorithm into a semi-supervised one. Results in the context of information retrieval will be presented.

(This is joint work with Katrin Kirchhoff)

Back to symposium main page