Résumé

This second campaign of the TREC Deep Learning Track was an opportunity for us to experiment with deep neural language models reranking techniques in a realistic use case. This year’s tasks were the same as the previous edition: (1) building a reranking system and (2) building an end-to-end retrieval system. Both tasks could be completed on both a document and a passage collection. In this paper, we describe how we coupled Anserini’s information retrieval toolkit with a BERT-based classifier to build a state-of-the-art end-to-end retrieval system. Our only submission which is based on a RoBERTa large pretrained model achieves for (1)a ncdg@10 of .6558 and .6295 for passages and documents respectively and for (2) a ndcg@10 of .6614 and .6404 for passages and documents respectively.

Détails

Actions

PDF