search

parthsarthi03 / raptor

The official implementation of RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval
https://arxiv.org/abs/2401.18059
MIT License
687 stars 98 forks source link

Clarification about the experiment setting in the paper #13

Closed kerkathy closed 3 months ago

kerkathy commented 3 months ago

Hi, thanks for the work and I found the idea really interesting!

Since I wanted to make a reproduction of the experiments, I have a small question. To my knowledge, your method is to build the tree first and query the tree later. However, in the QuALITY dataset, there is one passage of around 5000 tokens for each question. In this case, do you build one tree for each question using the passage corresponding to that specific question? Also, for the other datasets (NarrativeQA and QASPER), do you build a shared tree for all questions?

thanks again for your kind sharing!

parthsarthi03 commented 3 months ago

Hi! Thanks for your interest in the work! To clarify, the tree-building process is passage-specific but question-independent. For each dataset, we build one tree per passage. Since each passage has multiple associated questions, the same tree is used to answer all questions for that passage.

kerkathy commented 3 months ago

Thanks for your prompt and clear explanation 😄