Dataset Card for CaT-Bench

CaT-Bench is a benchmark dataset designed to evaluate large language models' (LLMs) understanding of causal and temporal dependencies in natural language plans, specifically in cooking recipes based on the English Recipe Flow Graph Corpus by Yamakata et al. (2020). It consists of questions that test whether one step must necessarily occur before or after another, requiring reasoning about preconditions, effects, and the overall structure of the plan.

Dataset Details

Dataset Description

CaT-Bench (Causal and Temporal Benchmark) is aimed at assessing the ability of language models to reason about causal and temporal relationships within natural language plans. The dataset is constructed from cooking recipes and contains 9,162 questions about causal dependencies spanning 300 unique plans. Each question asks whether a particular step in a recipe must occur before or after another step, challenging models to understand the underlying causal and temporal structure.

• Curated by: Yash Kumar Lal*, Vanya Cohen*, Nathanael Chambers, Niranjan Balasubramanian, Raymond Mooney (* equal contribution)
• Funded by: DARPA KAIROS program under agreement number FA8750-19-2-1003, National Science Foundation under award IIS #2007290, DARPA's Perceptually-enabled Task Guidance (PTG) program under Contract No. HR001122C007
• Language(s): English
• License: MIT License

Dataset Sources

• Repository: https://github.com/StonyBrookNLP/CaT-Bench
• Paper: https://arxiv.org/abs/2406.15823

Uses

Direct Use

CaT-Bench is intended for evaluating and benchmarking the performance of language models on tasks requiring causal and temporal reasoning within natural language plans. Researchers and practitioners can use this dataset to assess how well models understand dependencies between steps in a plan, such as preconditions and effects, and improve model architectures or training methods accordingly.

Dataset Structure

The dataset consists of:

• Plans: 300 unique cooking recipes.
• Questions: 9,162 binary (yes/no) questions about step dependencies.
• Annotations: Questions are labeled as dependent (steps are dependent) or non-dependent (steps are independent).

Dataset Creation

Curation Rationale

The dataset was created to address the need for evaluating language models' understanding of causal and temporal dependencies in natural language plans. While LLMs have shown impressive generative capabilities, their ability to comprehend and reason about the structure and dependencies within plans is less understood. CaT-Bench aims to fill this gap by providing a benchmark specifically focused on this aspect.

Source Data

Data Collection and Processing

The dataset is based on the English Recipe Flow Graph Corpus, which contains 300 English cooking recipes annotated with substep procedure dependencies. From this corpus:

• Selection: 300 recipes were selected for inclusion.
• Question Generation: For each ordered pair of steps, two binary questions were created regarding the necessity of one step occurring before or after another.
• Balancing: The dataset was balanced to have an equal number of dependent and non-dependent questions.
• Annotations: Steps were annotated based on whether there is a directed path between them in the recipe's dependency graph.

Who are the source data producers?

The source data originates from the English Recipe Flow Graph Corpus created by Yamakata et al. (2020). The recipes were collected from publicly available sources, and the dependency annotations were provided by the original authors.

Personal and Sensitive Information

The dataset does not contain any personal, sensitive, or private information. All data are based on publicly available cooking recipes and contain no personally identifiable information.

Citation

BibTeX:

Please cite both this benchmark and the original English Recipe Flow Graph Corpus dataset.

@inproceedings{lal2024catbenchbenchmarkinglanguagemodel,
    title={CaT-Bench: Benchmarking Language Model Understanding of Causal and Temporal Dependencies in Plans},
    author={Yash Kumar Lal and Vanya Cohen and Nathanael Chambers and Niranjan Balasubramanian and Raymond Mooney},
    booktitle={Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
    year={2024},
    url={https://arxiv.org/abs/2406.15823},
}

@inproceedings{yamakata-etal-2020-english,
    title = "English Recipe Flow Graph Corpus",
    author = "Yamakata, Yoko and Mori, Shinsuke and Carroll, John",
    booktitle = "Proceedings of the Twelfth Language Resources and Evaluation Conference",
    year = "2020",
    publisher = "European Language Resources Association",
    url = "https://aclanthology.org/2020.lrec-1.638",
    pages = "5187--5194",
    language = "English",
    ISBN = "979-10-95546-34-4",
}