Hugging Face's logo

Datasets:
henry12348
/
DiPlomat

Languages:
English
ArXiv:
License:
Dataset card Files Community
Dataset Viewer
Full Screen

The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider removing the loading script and relying on automated data support (you can use convert_to_parquet from the datasets library). If this is not possible, please open a discussion for direct help.

DiPlomat

Pragmatic reasoning plays a pivotal role in deciphering implicit meanings that frequently arise in real-life conversations and is essential for the development of communicative social agents. In this paper, we introduce a novel challenge, DiPlomat, aiming at benchmarking machines’ capabilities on pragmatic reasoning and situated conversational understanding. Compared with previous works that treat different figurative expressions (e.g. metaphor, sarcasm) as individual tasks, DiPlomat provides a cohesive framework towards general pragmatic understanding.

Dataset Details

The DiPlomat dataset owns 4,177 data and covers a vocabulary of 48,900 words. More than that, human-annotated answers reach an amount of 6,494, hold a vocabulary size of 20,000, and cover 5 types of reasoning. Along with the dataset, we propose two tasks:
Pragmatic Identification and Reasoning (PIR) and Conversational Question Answering (CQA). Furthermore, we provide the data that we use for zero-NLI.

Dataset Sources

Dataset Structure

Field Task
PIR_first Pragmatic Identification and Reasoning Subtask1
PIR_second Pragmatic Identification and Reasoning Subtask2
CQA Conversational Question Answering
NLI_with_context Zero-Shot NLI with context
NLI_without_context Zero-Shot NLI without context

NOTE: If you'd like to test on the whole PIR task, please don't change the order of PIR Subtask 1's and Subtask 2's test file's data, as both of them are deliberately arranged as the same order.

Dataset Creation

Source Data

We leverage the data of INTERVIEW dataset collected by Majumder et al as our source.

Annotating Process

Step I. Automatic Selection:

The extensive size of the source dataset introduces redundancy, and thus requires automatic measures to alleviate the burden of human annotation. Therefore, we employ algorithms and models to perform an initial filtering process.

Step II. Fine-grained Annotation:

We leverage Amazon Mechanical Turk to conduct detailed annotations of pragmatic turns within our dialogues. Workers participating in the annotation task are instructed to select all turns that exhibit a divergence between their literal meaning and their intended meaning. Due to the subjective nature of pragmatic reasoning, we request the workers to provide confidence scores along with reasons for their choices.

Step III. Human Refinement:

In this process, tasks for workers are formulated as multiple-choice questions. Previously collected human-annotated reasons are transformed into choices, utilizing a template format: [turn {turn_id}: {reason}]. In addition, to mitigate the impact of careless workers, we introduce a distractor choice for each gold choice.

Citation 

@inproceedings{li2023diplomat,
    title={DiPlomat: A Dialogue Dataset for Situated Pragmatic Reasoning},
    author={Hengli Li and Song-Chun Zhu and Zilong Zheng},
    booktitle={Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
    year={2023}
}

Dataset Card Contact

If there is any problem with the dataset, please email [lihengli@stu.pku.edu.cn](mailto: 2000017754@stu.pku.edu.cn).

Downloads last month
106