While BM25 did emerge from earlier work in the 1970s and 1980s (specifically building on the probabilistic ranking principle), I'm curious about your perspective on a few things:

What specific modern statistical approaches are you seeing as superior replacements for BM25 in practical applications? I'm particularly interested in how they handle edge cases like rare terms and document length normalization that BM25 was explicitly designed to address.

While I agree learning-based approaches have shown impressive results, could you elaborate on what you mean by search being "strictly dominated" by learning methods? Are you referring to specific benchmarks or real-world applications?

BM25 can be used as a starting point for a statistical learning model and more readily built on. A key advantage is that one gains a systematic way to reduce edge cases, instead of handling a couple, bc they’re so large as to be noticeable.