metadata

task_categories:
  - question-answering
  - table-question-answering
size_categories:
  - 1K<n<10K
language:
  - en
  - fr
configs:
  - config_name: queries_english
    data_files:
      - split: train
        path: english/train.csv
      - split: dev
        path: english/dev.csv
      - split: test
        path: english/test.csv
  - config_name: queries_french
    data_files:
      - split: train
        path: french/train.csv
      - split: dev
        path: french/dev.csv
      - split: test
        path: french/test.csv
  - config_name: corpus
    data_files:
      - split: english
        path: english/corpus.csv
      - split: french
        path: french/corpus.csv

Statcan Dialogue Dataset (Processed for Retrieval Tasks)

This is a variant of the Statcan Dialogue Dataset, which we processed specifically for multilingual retrieval (english, french). It contains everything in CSVs, rather than having metadata hosted separately.

Quickstart


from datasets import load_dataset

repo = 'McGill-NLP/statcan-dialogue-dataset-retrieval'
# load english queries, training split

queries_en = load_dataset(repo, 'queries_english', split='train')  # or 'dev', 'test'
queries_fr = load_dataset(repo, 'queries_french', split='train')
# Dataset({
#     features: ['query', 'query_id', 'doc_id'],
#     num_rows: 3782
# })

# load corpus (available in french and english)
corpus_en = load_dataset(repo, 'corpus', split='english')
corpus_fr = load_dataset(repo, 'corpus', split='french')
# Dataset({
#     features: ['doc_id', 'title', 'doc'],
#     num_rows: 5907
# })

The queries is given in list of dicts, with the following keys:

role: either "user" or "operator". Operator is the expert in charge of helping the user.
content: the utterance of the user or operator.

# ...continued
q = queries_en[0]['query']
doc_id = queries_en[0]['doc_id']
print("Query:", q)

# You can filter the corpus to find your table (english and french are the same docs)
d = corpus_fr.filter(lambda r: r['doc_id'] == doc_id)[0]
print(d['title'])
print('-'*50)
print(d['doc'])

Note that d['title'] only contains the title of the table, whereas d['doc'] is the full document and contains:

title,
time period,
table dimensions,
subject (for taxonomy),
name of survey where it was sourced,
update frequency,
columns for each of the dimensions

Warning: please do not do d['title'] + d['doc'] as that would be redundant.

License

By using this dataset, you agree to the the following terms of use and restrictions:

Terms of use

Researchers must agree to the following terms:

These data represent anonymized (de-identified) data from individuals. Best efforts have been implemented to ensure that all directly and indirectly identifiable information has been removed. Researchers who download this dataset must agree to notify Graeme Gilmour (graeme.gilmour <at> statcan.gc.ca) and Harm de Vries (harm.devries <at> servicenow.com) if any inadvertently remaining identifiable information is discovered during the process of re-using this dataset. Researchers must agree to destroy any version of this dataset containing identifiable information.
The terms of this dataset require that reusers give credit to the creators. It allows reusers to distribute, remix, adapt, and build upon the material in any medium or format, even for commercial purposes.
Have read and acknowledged the Appendix B (Dataset Card) of the latest version of the paper prior to using the dataset.

Restrictions

Downloaders cannot:

obtain information from the dataset that results in the researcher or any third party(ies) directly or indirectly identifying any participant with the aid of other information acquired elsewhere;
produce connections or links among or between the information included in the dataset and other third-party information that could be used to identify any individuals; and
extract information from the dataset that could aid researchers (downloaders) in gaining knowledge about or obtaining any means of contacting any individuals already known to the downloader/researcher