Engelmann, Björn: IRCologne at TREC 2021 News Track : Relation-Based Re-ranking for Background Linking

Abstract in English:

This paper presents our approach to the background linking task of the TREC 2021 News Track. The background linking task is to find a set of relevant articles in the Washington Post dataset containing helpful background information for a given news article. Our approach involved a two-stage retrieval process. In the first stage, the 200 most relevant documents were extracted from the entire corpus using BM25. The second stage involved re-ranking using similarity scores based on entities and relations extracted from the query document and the associated 200 relevant documents. For this task, we submitted five runs, each giving different weights to the entities and relations. Our best run received a nDCG@5 of 0.4423, and we were thus able to show that re-ranking with the use of relations leads to a slight improvement over the baseline without re-ranking.