README.md

metadata

dataset_info:
  features:
    - name: example_id
      dtype: int64
    - name: query
      dtype: string
    - name: query_id
      dtype: int64
    - name: product_id
      dtype: string
    - name: product_locale
      dtype: string
    - name: esci_label
      dtype: string
    - name: small_version
      dtype: int64
    - name: large_version
      dtype: int64
    - name: split
      dtype: string
    - name: product_title
      dtype: string
    - name: product_description
      dtype: string
    - name: product_bullet_point
      dtype: string
    - name: product_brand
      dtype: string
    - name: product_color
      dtype: string
    - name: source
      dtype: string
    - name: full_description
      dtype: string
    - name: Boost Product Index
      dtype: int64
    - name: description
      dtype: string
  splits:
    - name: train
      num_bytes: 130485155
      num_examples: 13985
  download_size: 58648377
  dataset_size: 130485155
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

This dataset is a sample for the Amazon Shopping Queries Dataset.

This dataset contains queries for which at least 10 products are available. The products if possible are exact matches to the query intent, or at least substitutes It was constructed as follows:

import pandas as pd

df_examples = pd.read_parquet("shopping_queries_dataset_examples.parquet")
df_products = pd.read_parquet("shopping_queries_dataset_products.parquet")
df_sources = pd.read_csv("shopping_queries_dataset_sources.csv")

df_examples_products = pd.merge(
    df_examples,
    df_products,
    how="left",
    left_on=["product_locale", "product_id"],
    right_on=["product_locale", "product_id"],
)

df_examples_products_source = pd.merge(
    df_examples_products,
    df_sources,
    how="left",
    left_on=["query_id"],
    right_on=["query_id"],
)

list_hits = []
for query_id in tqdm(list_query_id):
    df = retrieve_products(query_id, df_examples_products_source)
    list_len_desc = []
    for row_idx in range(len(df)):
        row = df.iloc[row_idx]
        full_description = format_product_details(row)
        list_len_desc.append(len(full_description))
    if len(df) >= 10:
        list_hits.append((df, np.mean(list_len_desc)))

# sort by length of full_description
list_hits = sorted(list_hits, key=lambda x: x[1], reverse=True)

df = pd.concat([x[0] for x in list_hits[:1000]])

The auxiliary functions are:

def format_product_details(product):
    template = "List of features:\n{features}\n\nDescription:\n{description}"
    features = product["product_bullet_point"]
    description = product["product_description"]
    return template.format(features=features, description=description)


def retrieve_products(query_id, df_examples_products_source):
    df = df_examples_products_source[
        df_examples_products_source["query_id"] == query_id
    ]
    # product_locale = en
    df = df[df["product_locale"] == "us"]
    # remove esci_label I
    df = df[df["esci_label"] != "I"]
    # remove product_description None
    df = df[df["product_description"].notnull()]
    # remove product_bullet_point None
    df = df[df["product_bullet_point"].notnull()]
    # if esci_label E > 10, use only those
    if df[df["esci_label"] == "E"].shape[0] > 10:
        df = df[df["esci_label"] == "E"]
    # if esci_label in [E, S ]> 10, use only those
    elif df[df["esci_label"].isin(["E", "S"])].shape[0] > 10:
        df = df[df["esci_label"].isin(["E", "S
    else:
        return []
    return df