WaterlooClarke at the TREC 2021 Conversational Assistant Track

This report details the generation of each of these runs and the outcome of these experiments. My overall approach can be explained as three steps: 1) query construction, 2) passage retrieval and ranking, and 3) passage re-ranking. The third step is optional and applies only to the clac*Rerank runs. Like everyone else participating this first year, my ability to explore alternatives for these steps was hampered by the lack of training data specific to this track, and so ultimately I adopted relatively simple methods for all steps. During an exploratory phrase, while these methods were being selected and developed, I did a fair amount of seat-of-the-pants judging and side-by-side comparison of results on selected training topics. I never want to read about physician assistants, Plessy v. Ferguson, or water molecules ever again. On the other hand, I have lots of ideas for improving these methods, and I’m looking for to the continuation of the track in TREC 2020.