CONFERENCE (INTERNATIONAL) TKB48 at TREC 2021 Fairness Ranking Track

Zhuoqi Jin (University of Tsukuba), Hideo Joho (University of Tsukuba), Sumio Fujita


November 15, 2021

Fairness ranking has been recently focused on, which aims to make ranking results fair while keeping relevant. The definition of fair- ness is diverse. TREC Fairness Ranking Track in 2021 took attention- weighted rank fairness (AWRF) [12] to fit the fairness aspect dis- tribution of ranking results to a population estimator pˆ reflecting the target distribution. TKB48’s approach was a post-processing method. We obtained an initial ranking using the BM25 score. We then set a bucket for each of 7 geographic areas in the dataset, and iterated the initial BM25 ranking to choose documents and put them into the bucket in a round-robin manner. As the track evaluated the top 20 results of the final ranking, the goal for us was to make the distribution of each area be the same as the target distribution in the top 20 results. We defined the individual fairness score so that we could choose whether a document should be put into the bucket by comparing an individual fairness score and BM25 score. The individual fairness score was based on how many documents in a certain area has been put into the final ranking. We chose one docu- ment with the highest combined score of fairness and relevance for one iterate turn of initial ranking. And we iterated 1000 times so that we could get a final ranking with 1000 documents. Our results ranked fifth out of 13 submissions on the TREC Fairness Ranking Track. Finally, we compared the results of different methods on the TREC Fairness Ranking Track and analyzed it.

Paper : TKB48 at TREC 2021 Fairness Ranking Track (external link)