search

microsoft / unilm

Large-scale Self-supervised Pre-training Across Tasks, Languages, and Modalities
https://aka.ms/GeneralAI
MIT License
20.19k stars 2.55k forks source link

HotpotQA on NLG task #189

Open ghost opened 4 years ago

ghost commented 4 years ago

Describe Model I am using (UniLM, MiniLM, LayoutLM ...):minilm

Is it a wise idea to use inverse HotPotQA and train minilm ckpt in s2s toolkit for the NQG task? Since hotpotQA generates multi-hop, will it make any sense? I was hoping to get more complex questions

donglixp commented 4 years ago

The pre-trained models might have learned some implicit multi-hop reasoning. The multi-hop paths can also be used as input to guide the generation. Moreover, the techniques proposed to improve multi-hop QA could be integrated into the inverse problem.

ghost commented 4 years ago

Thanks for the reply. I do see it has a great enhancement for Question generation. Since the model might have learned some multi-hop reasoning, is it advisable to train it on the inverse HotpotQA + Inverse Squad together? Also, for adding a multi-hot context, it will be better to provide context as only supporting facts or the entire paragraphs? (https://hotpotqa.github.io/explorer.html) I feel providing the entire paragraph will improve its multi-hop reasoning but might even get into longer sequence length issues. LMK

donglixp commented 4 years ago

It depends on whether we could extract accurate supporting facts. If the extracted facts are noisy, it would be better to feed the whole paragraph. Looking forward to your results.