Robertson & Zaragoza (2009) — the probabilistic relevance framework, and why nobody reads it

[akash-coded]

Reading Club. One paper per thread. The house position: replicate, do not cite.
A paper we can re-run against this corpus is worth far more argued about with numbers.

Paper: arXiv / FnTIR 3(4)

The claim, in one paragraph

Everyone uses BM25. Almost nobody has read where it comes from, which is why "why does term
frequency saturate?" stops most candidates in a depth round.

The claim: relevance ranking can be derived from the odds that a document is relevant, and BM25's
peculiar shape — a saturating tf/(k₁+tf), a length term interpolated by b, an IDF that goes
negative — falls out of that derivation rather than being tuned into existence.

How we would test it here

notebooks/04 §4.2–4.3 builds the inverted index by hand, and
L03 makes you compute the IDF and find where it
crosses zero. Sweep k₁ and b on this corpus and see whether the shape the paper predicts is
the shape you get.

Two or three things to argue about

1. Is the 2-Poisson eliteness story load-bearing, or a post-hoc justification for a curve that worked? The saturating form is an approximation to a model with no cheap closed form. Would a different concave bounded function do as well?
2. IDF goes negative above ~50% document frequency. Lucene's classic default floors it at zero. Which is right, and is the disagreement empirical or philosophical?
3. Term independence is false — 'machine' and 'learning' are not independent. Why does BM25 work anyway, and what does that tell you about how much rigour a ranking function actually needs?

---

If you have run any part of this, post the numbers — a replication with an interval outranks
every opinion in the thread, including mine.