JDivPS

JDivPS: A Diversified Product Search Dataset

Dataset Download

As the files are too large, we provide the following approaches to get the data:

• For international users, you can use Google Drive or JD JoyBox-HK to download the file. Notice that JoyBox-HK requires an E-mail address to log in.

• For Chinese users, you can use JD Joybox to get the file. The access password is lfe39l. You need an account on JD.com to access the data.

If you are facing difficulties accessing the data, feel free to contact us at the following E-mail address:

• [email protected]
• [email protected]
• [email protected]

Dataset Structures

The dataset includes the following 5 files:

data_release
  ├── dict_product_text_release.pkl.gz
  ├── product_uvctr_dict_release.pkl.gz
  ├── query_intent_label_tr.csv
  ├── query_intent_label_ts.csv
  ├── query_suggestions_release.pkl.gz
  └── query_product_features_release.pkl.gz

The pkl.gz files are compressed binary files that can be opened in Python with the pickle and gzip packages. More details can be found in data_release/check_data.py. All those text contents are tokenized into integer ids with a private tokenizer. The content descriptions of the files are listed as follows:

• dict_product_text_release.pkl.gz: the text metadata of the products. It is a Python dictionary with the following structure:

  {product_id:[product_name,category_name,brand_name,size,attribute,color]}
  

  Attribute	Description
  product_id	the product’s anonymized id
  product_name	the product's anonymized term ids
  category_name	the product category's anonymized term ids
  brand_name	the product brand's anonymized term ids
  size	the product size's anonymized term ids
  attribute	the product attribute's anonymized term ids
  color	the product color's anonymized term ids

It should be addressed that the size, attribute and color of the product may be empty.

• query_suggestions_release.pkl.gz: the query suggestions corresponding to the query. It is a Python dictionary with the following structure:

  {query:[suggestions1, suggestions2, ...]}
  

• product_uvctr_dict_release.pkl.gz: the popularity features of the products. It is a Python dictionary with the following structure:

  {product_id:[uv,pv,ctr]}
  

  Attribute	Description
  product_id	the product’s anonymized id
  uv, pv, ctr	UV, PV, and CTR score of the product

• query_product_features_release.pkl.gz: all the features of every existing query-product pair. Details about the features can be found in our paper. It is a Python dictionary with the following structure:

  {(query,product_id):[relevance_score,tf_idf_title,tf_idf_category,tf_idf_brand, bm25_title,bm25_category,bm25_brand,uv,pv,ctr]}
  

  Attribute	Description
  query	the query's anonymized term ids
  product_id	the product’s anonymized id
  relevance_score	relevance of the product to the query
  tf_idf_title	tf-idf score of the product's title
  tf_idf_category	tf-idf score of the product's category
  tf_idf_brand	tf-idf score of the product's brand
  bm25_title	BM25 score of the product's title
  bm25_category	BM25 score of the product's category
  bm25_brand	BM25 score of the product's brand
  uv,pv,ctr	UV, PV, and CTR score of the product

  Notice that the UV, PV, and CTR are identical to the features in product_uvctr_dict_release.pkl.gz.

The initial ranking lists with relevance scores can be generated with the data_release/generate_initial_ranking_list.py file.

The Structure of Intent Annotations

• query_intent_label_tr.csv: the query intent annotations for the training set.
• query_intent_label_ts.csv: the query intent annotations for the test set. All those csv files are separated by \t in the following format:

query\t intent\t product_id\t label

Attribute	Description
query	the query's anonymized term ids
intent	the anonymized term ids of a user intent
product_id	anonymized product_id of a product in the initial product list
relation (0/1)	relevance of a product to the intent

Notice that we only provide the positive annotations to reduce the size of the file. Here query and intent are the tokenized integers concatenated with , as a separator.

Evaluation

The diversity measures can be evaluated with the official TREC ndeval tool. More details can be found in TREC official sites.

The Pretrained BERT model

We provide two 12-layer BERT models with the same word tables as our private tokenizer.

Pretrained Model

The special token ID map of our private tokenizer is listed as follows:

[UNK] 1
[SEP] 3
[PAD] 0
[CLS] 2
[MASK] 4

We provide a BERT model denoted as scratch_bert which is pretrained on over 10M product titles with the task of Masked Language Model(MLM):

• Download path: Google Drive, JD JoyBox-HK, JD JoyBox password: yrh7z5

It can be loaded and used with the BertModel.from_pretrained method of Huggingface Transformers.

Fine-tuned Model

Based on scratch_bert, we provide another model denoted as rel_bert for computing the relevance between queries and product titles. We use the relevance model in the platform as a teacher model to distill rel_bert for computing the relevance between a query and a product title.

• Download path: Google Drive, JD JoyBox-HK, JD JoyBox password: ud9oeq

More details can be found in load_relevance_model.py for the instructions for loading the relevance model checkpoint.

We will later release more pretrained and fine-tuned model checkpoints to support the research based on JDivPS dataset.

License

This repository is licensed under Apache-2.0 License.

The JDivPS dataset is licensed under CC BY-NC-SA 4.0.