LUCENE-6818 we have DFISimilarity which implements normalized chi-squared distance.
But there are other alternatives (as described in http://trec.nist.gov/pubs/trec21/papers/irra.web.nb.pdf):
- normalized chi-squared: "can be used for tasks that require high precision, against both short and long queries"
- standardized: "good at tasks that require high recall and high precision, especially against short queries composed of a few words as in the case of Internet searches"
- saturated: "for tasks that require high recall against long queries"
I think we should just provide the three independence measures, and let the user choose. Similar to how we do DFR/IB/etc.