← smac.pub home

Contextualized PACRR for Complex Answer Retrieval

pdf bibtex slides poster system description paper non-refereed

See revised version, published in SIGIR 2018 link

Authors: Sean MacAvaney, Andrew Yates, Kai Hui

Appeared in: Proceedings of the 26th Text REtrieval Conference (TREC 2017)

Links/IDs:
DBLP conf/trec/MacAvaneyYH17 Google Scholar 7wWfoDgAAAAJ:2osOgNQ5qMEC Semantic Scholar 0ed5d050c1a983bba0f3eb57cccb5701d9d5de1d smac.pub trec2017-car

Abstract:

In this work, we present our submission to the TREC Complex Answer Retrieval (CAR) task. Our approach uses a variation of the Position-Aware Convolutional Recurrent Relevance Matching (PACRR) deep neural model to re-rank passages. Modifications include an expanded convolutional kernel size, and contextual vectors to capture heading type (e.g. title), heading frequency, and query term occurrence frequency. We submitted three runs for human relevance judgments by TREC varying which contextual vectors are included, and the number of negative samples used when training. Our approach yields a MAP of 0.241, R-Prec of 0.321, and MRR of 0.520 when using the term occurrence frequency run in the lenient evaluation environment.

BibTeX @inproceedings{macavaney:trec2017-car, author = {MacAvaney, Sean and Yates, Andrew and Hui, Kai}, title = {Contextualized PACRR for Complex Answer Retrieval}, booktitle = {Proceedings of the 26th Text REtrieval Conference}, year = {2017} }