Check out all the details and more findings here:
arxiv.org/abs/2503.22508

I will also present this at the #IR4GOOD at #ECIR2025 next week, alongside many other contributions from
@irglasgow.bsky.social. Looking forward to continuing these discussions at #SIGIR2025.

Then, we try to zero-shot these fine-tuned models on other language pairs. Some are related to French-Catalan, such as Occitan, and some are entirely unrelated, such as Mandarin.

We see it does transfer to Occitan-French pairs and in Cantonese-Mandarin pairs (more on the paper)

We fine-tuned them on Catalan queries and French Docs and see that we can regularise the models to be more robust on Catalan (and we see some gains in French!)

What will happen if you fine-tune neural rankers such as BGE-M3 and ColBERT-XM on low-resource queries and high-resource documents of two different (albeit related) languages?

Will it help regularize the rankers on these similarities?

We translated five collections of mMARCO into similar languages and evaluated retrieval methods based on how well they would perform if the queries were expressed in a similar low-resource language.

It turns out they do not perform very well (an understatement).