Closed GabrielBellard closed 4 years ago
Out of the box! The model preprocesses/tokenizes by itself.
This repo / paper has info for tuning BM25, with anserini, and comparable experiments with tuned ES https://github.com/nyu-dl/dl4marco-bert
I seems like it’s about a max 0.04 MRR boost
In your experiments, did you tune ElasticSearch with analyzers (stem, lemmatization, stopwords...) or you only used ES out of the box?