DCH-2: Training Data for the DialEval-2 Task

For detail of the NTCIR-16 Dialogue Evaluation Task (DialEval-2), see here.

Overview of the DCH-2 Dataset

Training Data

The Chinese training dataset contains 4,390 (4,090 for training + 300 for dev) customer-helpdesk dialgoues which are crawled from Weibo. All of these dialogues are annotated by 19 annotators.

In the DCH-2 dataset, as we finished the translation of all the Chinese dialogues, the English dataset contains 4,390 dialogues as well. All the English dialogues are manually translated from the Chinese dataset. The English dataset shares the same annotations with the Chinese dataset.

Test Data

Will be released in Dec 2021 according to the task schedule.

Annotation

We hired 19 Chinese students to annotate the training/dev dataset in 2018. In 2019, the test dataset of DialEval-1 were annotated by another group of annotators. Thus, there may be a gap between the training data and test data, as the dialogue annotation is quite subjective.

Format of the JSON file

Each file is in JSON format with UTF-8 encoding.

Following are the top-level fields:

Each element of the turns field contains the following fields:

Each element of annotations contains the following fields:

Nugget Types

drawing

Dialogue Quality

Scale: [2, 1, 0, -1, -2]

Note

To protect the privacy, some sensitive information in the dialogue data has been masked. For example, we replaced telephone numbers with 123456789, and email addresses with XXX@YYY.com

Access to the dataset

To obtain the dataset, please fill and submit our user agreement form.

Conditions and Terms

See here.

Have questions?

Please contact: dialeval2org@list.waseda.jp

Links