Question about alignment of passages and queries
Hello, I saw that in your paper, you mentioned: "both the queries and the relevant passages are aligned across languages". However, in this dataset, I noticed the queries are not fully paralleled across all languages. For instance, there are only 3 queries written in Vietnamese. Similarly, in the document file, I didn't find the translation of some English documents in other languages.
So I wanna ask if I mistakenly understand anything or is it a mistake in the dataset you uploaded?
Thanks in advance!
Hi @JRQi , thank you for your interest in our work! I double-checked your concern, and the currently uploaded dataset is correct. In our paper, what we meant by "both the queries and the relevant passages are aligned across languages", is that for for a given territory, we will have the queries and passages aligned across the relevant entities. So for Vietnamese, it is correct that there are only 3 queries, since the country of Vietnam only has 3 disputed territories in the original dataset. We see how the wording in that quote is a bit confusing, so we will revise it to be clearer in the next revision.
As a side note, the paper only covers the following modes:
dsoa_vi = load_dataset("borderlines/bordirlines", "vi", split="openai.qlang")
dsm3_vi = load_dataset("borderlines/bordirlines", "vi", split="m3.qlang")
Since that, we have added 3 additional modes, for a total of 8 IR sets, which you can load by:dsd_vi = load_dataset("borderlines/bordirlines", "vi")
As for "in the document file, I didn't find the translation of some English documents in other languages", note that none of the documents are translated. Instead, we use different language versions of Wikipedia. So the passages for the country "France" in {en, fr, ...} are all distinct, as for any country or territory. I hope that clears things up!
PS: We have also seen your great work on Cross-lingual consistency, and have considered adapting the ranking-based approach for evaluation of LLM outputs. Perhaps we can further discuss over email?