malllabiisc / EmbedKGQA

ACL 2020: Improving Multi-hop Question Answering over Knowledge Graphs using Knowledge Base Embeddings
Apache License 2.0
412 stars 95 forks source link

[MetaQA] Training data contains knowledge base triplets as questions #120

Closed AndreiCComan closed 2 years ago

AndreiCComan commented 2 years ago

Hello,

thank you for your valuable and interesting work!

I was wondering if there is any specific reason on why the knowledge base triplets have been inserted within the training data as questions that follow the same format as the original MetaQA dataset questions. Have those triplets been used during the training phase or should I refer only to those files that have the suffix "_old"?

Best, Andrei

apoorvumang commented 2 years ago

Hi, thanks for your interest

We found that adding the KG triplets as 'qa pairs' during QA training improves performance slightly (compared to using KG triplets just during KGE training). You can use both kinds of files, those with and without the suffix "_old", but the ones with QA pairs should be better performing.

AndreiCComan commented 2 years ago

Hello,

thank you for your prompt reply and for clarifying the matter.

Best, Andrei