akariasai / cora

This is the official implementation of NeurIPS 2021 "One Question Answering Model for Many Languages with Cross-lingual Dense Passage Retrieval".

Hi there 👋

I am a fifth-year PhD student at Paul G. Allen School of Computer Science & Engineering, University of Washington advised by ‪Hannaneh Hajishirzi‬. I work on Natural Language Processing as part of UW NLP, with a focus on retrieval-augmented language models. Before UW, I obtained a B.E. degree in Electrical Engineering and Computer Science (EEIC) from The University of Tokyo.

Contact

• Personal Website: akariasai.github.io/
• Twitter: @AkariAsai
• email: my_first_name[at]cs.washington.edu

cora's Issues

Add requirements

Could you please provide the full requirements in requirement.txt for easier running your code?

Adding the language id twice to the question before passing it to mGEN

Hi,

Thanks for uploading CORA on github! I am trying to use your package in my project, and wanted to make sure if it was the intention of the authors to add the language_id twice to the outputs of the mDPR before passing it to mGEN.

In mDPR/dense_retriever.py, in the method parse_qa_jsonlines_file, the 2 - letter language id is added to the question while encoding the question for mDPR. Considering this was intentional, the 2 letter language id is added again while converting mDPR outputs to seq2seq (here)

What ends up happening is that before the input sequence is sent into mGEN, the question ID is appended in the end by the language id twice, both being the same. We would follow the same format if the authors intended it to be so.

Unable to reproduce the results in CORA paper

Hi authors! Thanks for your great work on the CORA system. I tried to reproduce your work, however, when I follow the instructions provided in README, the evaluation results are much lower than what you presented in CORA paper. Are there any potential reasons for this?

Hardware requirement to build a FAISS index from the dense embedding?

Hi,

First of all, thank you for sharing the great work!

I'm trying to replicate the implementation on another platform (ParlAI) and I have been building a FAISS index from the multilingual embeddings provided in this repo on a GPU with 50 GB vram. I used the FAISS compression techniques so it's supposed to build an index much faster. However, the program has been running for ~48 hours now and not stopping. I've also tried building it without the compression techniques but that took days and the cost became to high for me.

I'm wondering what is the hardware requirement to build such an index and how long would it take? Also would it be possible to share the index file you built?

Thank you so much!