Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

feature request: hits (or accuracy?)

See original GitHub issue

Hi,

@osf9018 mentioned it in #2 but I guess it’s better to create a specific issue.

Motivation

It is often difficult to estimate the total number of relevant document for a query. For example, in Question Answering, if you have a large enough Knowledge Base, you can find the answer to your question in a surprisingly large number of documents that one cannot annotate in advance. Because of this, the relevance of the document is often estimated on-the-go, by checking whether the answer string is in the document retrieved by the system.

Because of this, recall is not an appropriate metric. However, one way to circumvent this is to compute recall “as if” there was only a single relevant document. After averaging over the whole dataset, it corresponds to the proportion of question for which the system retrieved at least one relevant document in top-K. This is what @osf9018 and I call “hits@K” (I can’t remember but I’ve seen it in a paper) and others, such as Karpukhin et al., call “accuracy”. Accuracy is a confusing term IMO.

The request

Would you be interested in implementing or integrating this feature in your library? It might take some renaming but it could be implemented very easily by using the _hits function. It is simply min(1, _hits(qrels, run, k))