J-MADRAL/AmazonESCI

Amazon ESCI is a large-scale test collection for product search, comprising:
(i) a corpus of 1.2M products of the english portion of Amazon ESCI
(ii) a set of queries with 74.9k, 13.5k, and 8.9k queries in the training, validation, and test sets, respectively.
The relevance judgements comprise 20.15 assessments per query on average, divided in four relevance labels: Exact, Substitute, Complementary, and Irrelevant, with gains of 1.0, 0.1, 0.01, and 0.0 associated. Only Exact products are treated as positive products.

This dataset is composed of the following files:

• catalog.jsonl: It contains the data for each product in the corpus, each line encoding one product in JSON format. Each record has the following data:
  • id: The unique identifier of the product.
  • title: The title of the product.
  • description: The description of the product.
  • bullet_point: The concatenation of the bullet points information for the product.
  • categories: The list of the categories hierarchy (up to 4-th level) associated to the product.
  • details: Additional key-value metadata associated to the product.
  • image: URL to a medium-sized image of the product. "" if the data is not available.
  • brand: The brand metadata associated with the product. "" if the data is not available.
  • color: The color metadata associated with the product. "" if the data is not available.
  • price: The integer value of the product price (i.e., €39.99 => 3999). -1 if the data is not available.
  • num_ratings: The number of ratings reported on the Amazon page of the product. -1 if the data is not available.
  • avg_ratings: The average of ratings reported on the Amazon page of the product. -1.0 if the data is not available.
  • template: The page template used by Amazon for the product webpage. "" if the data is not available.
  • labels: Key-value dictionary containing the aspect number associated with the corresponding label. -100 if the label is not available for that aspect.
  • aspects: List of aspect labels, as used in J-MADRAL. -100 if the label is not available for that aspect.
• catalog-ids.txt: It contains the ID of every product, one per line. With catalog-offsets.txt, it enables random access to the products in the catalog.
• catalog-offsets.txt: It contains the number of bytes occupied by each product, one per line. The data of the k-th product is found in catalog.jsonl from start=sum(offsets[:k]) and end=sum(offsets[:k+1]) byte offsets. With catalog-ids.txt, it enables random access to the products in the catalog.
• <aspect>-mapping.json: The mapping between aspect names and the associated numerical value, provided in JSON format.
• <train/valid/test>_qrels.qrels: The relevance judgements using TREC format for the training, validation, and test sets, respectively.
• <train/valid/test>_queries.jsonl: It contains the data for each query in the dataset, each line encoding one query in JSON format, for the training, validation, and test sets, respectively. Each record has the following data:
  • id: The unique identifier of the query.
  • text: The textual content of the query.
  • labels: Key-value dictionary containing the aspect number associated with the corresponding label. -100 if the label is not available for that aspect.
  • aspects: List of aspect labels, as used in J-MADRAL. -100 if the label is not available for that aspect.