Test Dataset has no positive documents
#4
by
nv-bschifferer
- opened
Hello,
the paper Unifying Multimodal Retrieval via Document Screenshot Embedding shares results based on evaluating NQ. The NQ test split has no positive documents in the list. Is this an error? How did you evaluate the test data?
Hi @nv-bschifferer , thanks for your interest in our work.
Similar to DPR work, we use answer exactly match to evaluate whether a retrieved document is positive.
for a retrieved Wikipedia screenshot, we check if the answer string is in the text mirror of the screenshot.
please check https://github.com/texttron/tevatron/blob/e75b70ffa05c62dba5754159f5328613a63ad47b/examples/dse/eval_retrieval.py#L229 on how we evaluate using the wiki-ss-nq test split.