[Dmbu-l] Data Seminar

Harshal Chaudhari harshal at bu.edu
Tue Nov 1 11:34:51 EDT 2016


*Data Seminar*
*Proximity and HIT Based Candidate Generation for Cascading Rankers*
*Benedetto Buratti, NYU*
*Friday, November 4, 2016 at 10:30am in Hariri Seminar Room*

*Speaker: *Benedetto Buratti, NYU

Benedetto is a prospective student visiting our department this Friday.
Currently, he is a visiting researcher at  Tandon School of Engineering,
NYU, working on efficiency issues in web search engines. He is also the
co-founder and CEO of two startups, Bloomia and MEMio. His talk provides us
an opportunity to learn about his current work and interact with him,
possibly answering questions that he may have regarding BU and life in
Boston. Hope to see you guys there!

*Title: *Proximity and HIT Based Candidate Generation for Cascading Rankers

*Abstract: *Complex ranking functions are able to improve search engines
effectiveness but are known to entail remarkable efficiency issues. To face
those challenges modern search engines have adopted a coarse-to-fine
evaluation model known as cascading rankers. This model consist of a
layered sequence of pruning and ranking functions, that selectively
processing query relevant documents, guarantee high effectiveness and
performances.

A recent work by Wang et al. showed that using an additional indexing
layer, based on a query distribution and a language model, is possible to
efficiently generate high quality candidates for the further layers in the
cascade. In this paper we expand their framework including term’s proximity
and query independent scores. Using a quality and language model we have
built two indexing structures, one for proximity and one for HIT score. We
have used those structures to build an additional ranking layer to quickly
retrieve promising documents showing that those features are crucial to
increase result quality even further.

We have provided an extensive end to end comparison with the state-of-the
art learning to rank algorithms showing that our candidate generator
achieve a comparable retrieval effectiveness in a fraction of the time

--
Harshal Chaudhari
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://cs-mailman.bu.edu/pipermail/dmbu-l/attachments/20161101/b0a6aea4/attachment.html>


More information about the Dmbu-l mailing list