Use DirichletLM for retrieval
Based on results in https://webis.de/publications.html#filter:Assessing%20Argument%20Relevance
Not sure if we need to go Terrier for that? Anserini might also be an option if we want to stay with Lucene/Elastic (have not checked whether they support it). I have not used Terrier so far.
Switching out Lucene would be a good test to see how general the framework is ;) (in theory it should work without much problems...)