Page MenuHomePhabricator

Framework for checking sources on Wikidata (Does the source actually say what we claim it says?)
Open, MediumPublic

Description

Wikidata has a lot of references. It'd be great to have tools to allow checking if those references actually say what we claim and flag problematic references.

The work in T156389 might be relevant.

Event Timeline

Lydia_Pintscher raised the priority of this task from to Medium.
Lydia_Pintscher updated the task description. (Show Details)
Lydia_Pintscher renamed this task from framework for source checking to Framework for checking sources on Wikidata (Does the source actually say what we claim it says?).Apr 7 2018, 11:27 AM
Lydia_Pintscher updated the task description. (Show Details)
Lydia_Pintscher moved this task from Backlog to Project on the Wikimedia-Hackathon-2018 board.

@maiarocg: Hi and welcome! You claimed this task a while ago. How is it going? Do you need any help with anything?

Hi all

This problem is called Natural Language Inference (NLI) also known as textual entitlement . It is a super hot problem now in the NLP community, but imho research is still far away from producing usable tools in the Wikipedia context. This also requires a lot of computational resources (GPUs) to train.

Anyhow, I'm exploring if would be possible to create a usable API where you could send a claim and a document and the API will tell the relation between those pieces (confirm, reject, no information). I think the algorithm won't work well with subtle issues (eg. the references is talking about the main topic of the item, but does not content the specific information about the claim), but could be able to catch if the document (reference) is completely unrelated.

I'll keep you updated.