Iterative retrieval in case of non-unique top-k retrieval

[dhdhagar]

Hi! Thanks for this amazing work, and for making your code open-source.

I'm trying to figure out where in the code is non-unique passage retrieval handled that ensures that the final k results are unique. According to this footnote on page 3 in your paper "Phrase Retrieval Learns Passage Retrieval, Too", it seems that you perform iterative retrieval to achieve this. Could you point me to the code where this is happening?

[jhyuklee]

Hi @dhdhagar! Thanks for your issue.

In this makefile line:

DensePhrases/Makefile

Line 490 in 3a0365e

--top_k 200 \

you can see that we are retrieving top 200 phrases first. And the aggregation based on the unique passages happen here:

https://github.com/princeton-nlp/DensePhrases/blob/main/densephrases/index.py#L430

For our current datasets provided, this is enough to ensure outputting 100 passages per query. There are a very small number of edge cases where top 100 passages are not retrieved even with this setting, and you can enlarge the top-k to 400 in the makefile. Currently there's no automatic procedure for this.

[dhdhagar]

That makes it clear, thank you! So, --top_k is used to fetch a larger number of phrases, and the length of the final results depends on --psg_top_k.