Ad-hoc Retrieval

Given an arbitrary query, retrieve tok-\(k\) relevant document from the collection. For the image suggestion task, the query is a text, where the document is an image. (For the image promotion task, the query is an image, where the document is a text.) See about for more details of these two tasks.

  • A text is composed of textual information, e.g., title and context decriptions.
  • An image is composed of its visual information (pixel values) and metadata (reference, alt-text, and attribution captions).

Submission Format

Submissions for the ad hoc retrieval task should be in standard TREC 6-column format.

Query_ID Q0 DOC_ID RANK SCORE RUN_ID
  • Query_ID (DOC_ID): the unique identifier for each text (or image)
  • Q0: placeholder
  • RANK: the rank of the retrieved DOC_ID (for Query_ID)
  • SCORE: the score of the retrieved DOC_ID (for Query_ID)
  • RUN_ID: systems’s identifier

Example (Text-to-Image) Run:

wit-test-topic-000000001 Q0 ba65386c-63de-3471-8473-985f6a102607 1 69.9 boring_system
wit-test-topic-000000001 Q0 4c8c424a-149d-342d-820e-6e5de6df6008 2 65.3 boring_system
...
wit-test-topic-000000001 Q0 a26a103e-bc5c-3f25-b620-60af7ace8257 1000 1.4 boring_system
wit-test-topic-000000002 Q0 4c8c424a-149d-342d-820e-6e5de6df6008 1 71.5 boring_system
...

Participants are asked to return a ranked list of at most \(k=1,000\) documents for each query.

Evaluation

Example Usage (MRR@10, Recall@1000, success@1000):

trec_eval -c -M 10 -m recip_rank <path-to-qrels> <path-to-runfile>
trec_eval -c -m recall.1000 success.1000 <path-to-qrels> <path-to-runfile>

Development Dataset