COLD: Causal reasOning in cLosed Daily activities

Summary

This paper proposes the COLD (Causal reasOning in cLosed Daily activities) framework, aiming to bridge the gap between open-ended causal reasoning and symbolic representation-based question answering. The framework leverages human understanding of daily real-world activities to reason about the causal nature of events. The authors create a large set of causal queries and evaluate multiple Large Language Models (LLMs) on these queries. The findings show that causal reasoning is challenging for LLMs, even for activities considered trivial for humans. The authors also explore the causal reasoning abilities of LLMs using the backdoor criterion. The key contributions of this work are the development of the COLD framework, the creation of a substantial number of causal queries, and the evaluation of LLMs' performance on causal reasoning tasks. The findings highlight the need for further analysis using real-world events to properly validate LLMs' understanding of causality.

Strengths

- The COLD framework effectively bridges the gap between open-ended causal reasoning and symbolic representation-based question answering, utilizing human understanding of daily activities as a solid foundation. - The paper addresses the crucial issue of evaluating LLMs' causal reasoning capabilities, emphasizing the significance of investigating and validating their intellectual capabilities. - The paper is well-written, with clear explanations, logical flow, and concise language, ensuring effective communication of key points. - The evaluation of multiple LLMs on a large set of causal queries reveals limitations in their causal reasoning abilities, while the exploration using the backdoor criterion provides valuable insights into causal strength between events.

Weaknesses

Generally, I believe this paper makes good contributions. However, there are some minor issues that need to be addressed: - Since the queries are mostly automatically generated, it is necessary to support them with human annotations or expert evaluations in order to confirm the reliability of the generated queries. - The observational graphs are created through human annotations, which limits their capacity to cover a wide range of concepts. It would be better to discuss automated approaches for constructing such graphs in order to facilitate large-scale causal benchmarking. - Can the synthesized queries be used for fine-tuning? I am interested in whether splitting the observation graphs into different sets and training them on queries synthesized from the training graphs would significantly improve performance. This could greatly enhance the comprehensiveness of the paper. - Additionally, another set of baselines focusing on zero-shot commonsense question answering should be evaluated as well. It would be interesting to see whether transformations from commonsense knowledge bases can benefit causal reasoning tasks. I recommend checking these papers for reference. - Ma, K., Ilievski, F., Francis, J., Bisk, Y., Nyberg, E., & Oltramari, A. (2021, May). Knowledge-driven data construction for zero-shot evaluation in commonsense question answering. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 15, pp. 13507-13515). - Wang, W., Fang, T., Ding, W., Xu, B., Liu, X., Song, Y., & Bosselut, A. (2023, December). CAR: Conceptualization-Augmented Reasoner for Zero-Shot Commonsense Question Answering. In Findings of the Association for Computational Linguistics: EMNLP 2023 (pp. 13520-13545). - Kim, Y. J., Kwak, B. W., Kim, Y., Amplayo, R. K., Hwang, S. W., & Yeo, J. (2022, July). Modularized Transfer Learning with Multiple Knowledge Graphs for Zero-shot Commonsense Reasoning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (pp. 2244-2257). - Lastly, there are some grammar typos that need to be corrected. For example, the caption of table 3 should not have "Language" capitalized.

Questions

Please refer to the weakness section for my questions.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have dedicated a section explicitly discussing the limitations of their proposed COLD framework and offering potential solutions. The in-depth discussion contributes significantly to the paper, enhancing its overall quality.

Summary

This paper proposes causal reasoning in closed daily activities, which combines the causal reasoning works of open-ended causal reasoning via causal commonsense reasoning and symbolic representation-based question answering for theoretically backed-up analysis. By creating a dataset containing about 8 million queries, it can estimate the causal reasoning performance of pretrained language models. Moreover, the paper introduces theroies of causal inference (such as backdoor adjustment) to conduct in-depth analyses of causal reasoning in pretrained language model.

Strengths

1. The number of constructed dataset samples is large. 2. The usages of ATE and backdoor adjustment are novel 3. The analyses of causal reasoning in pretrained language models are thorough and in-depth.

Weaknesses

1. The paper is not well demonstrated, readers might be confused when reading sequentially 2. Just for the dataset itself, there is no differences between cold and copa. Even copa may have a better quality. 3. The dataset is limited to six scenarios, many of them are just paraphrases of the original ones. 4. Lack of human performance as a reference.

Questions

1. In lines 78-80, you give a event pair, which could be erroneously treated causal-related by humans. What is the differences between this event pair and the correct cause-effect pairs in your dataset? Most of the causal relations in your dataset are about causation of necessarity. 2. Even your dataset is in closed daily activities, the final version of the queries do not reflect this. 3. Why is the definit setting (Cold) bettern than the plausible setting (COPA)? I think causility is essentially a question of plausibility, there is no absolute causal relationship between two events. 4. How would humans perform on this dataset? Can humans obtain this answer by just look into the temporal relationship between events? How many queries can be answered correctly from a statistical perspective (not from the perspective of LMs)?

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

None

Summary

The paper proposes a dataset for evaluating the causal reasoning capabilities of LLMs by grounding evaluation in human understanding of real-world daily activities. The authors address the gap between open-ended causal commonsense reasoning and symbolic question answering by introducing the COLD framework, which generates causal queries based on daily activities. The paper describes the creation of causal graphs through crowd-sourcing. The authors test different LLMs using these causal queries, and show that even simple causal reasoning tasks remain challenging.

Strengths

1. Its approach to evaluating causal reasoning in LLMs by grounding the evaluation in real-world activities. 2. The use of a large dataset to create extensive causal queries. 3. The detailed experimentation with various open-source LLMs and the plan to publicly release the framework and datasets.

Weaknesses

1. The data used in the study appears overly simplistic, focusing on basic daily activities which may not provide a robust test for causal reasoning. If the goal is to evaluate how well models understand causal reasoning I would stay close to how we perform causal inference in science (i.e. gather data, understand whether we have an identification strategy, compute treatment effects/learn causal graph). 2. Causal Commonsense Reasoning is not causal in the sense of statistical causality. Almost all uses of causality in science are either: 1. To estimate a treatment effect in the real-world 2. To discover a causal graph, again in the real-world. This is done either from interventional or from observational data. This data doesn’t capture any of this, so I wonder what the usefulness of it is in relation to the causality literature. 3. The insights drawn from the experiments are not clearly articulated. What exactly has been learned from this new extensive dataset and all of the experiments? The ATE experiments are particularly hard to understand. ATE measures the marginal effect of a treatment on an outcome. Comparatively, this dataset includes binary questions about causal triplets. In the results the authors show accuracy measures, I don’t understand how to interpret this as the ability to properly estimate the effect, and I wasn’t able to understand how to do so from the authors’ description. 4. The examples provided (e.g. Figure 3) often lack clarity, as both options seem plausible effects without additional context or a causal graph.

Questions

1. What specific insights have you gained from your experiments with this dataset? Can you elaborate on how these insights contribute to our understanding of the causal reasoning capabilities of LLMs? Could we use it to propose any improvements to the current models? 2. Are there plans to incorporate more traditional causal inference use-cases, such as treatment effect estimation or causal graph discovery, into your framework? I mean this in the sense of evaluating the capabilities of LLM in performing these tasks. 3. In the first example in Figure 3, both options appear plausible without additional context. Can you provide a detailed explanation of how the correct choice is determined in such cases, and whether there is a mechanism to ensure that the queries are unambiguous?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The authors acknowledge limitations of their work, such as the restricted set of activities and the challenge of creating causal graphs for more complex long-term tasks.

Summary

This paper presents a new causal reasoning dataset for LLMs. The main motivation of the dataset is to bridge the gap between casual commonsense reasoning datasets and symbolic representation-based causal reasoning datasets. Specifically, the proposed dataset collects crowd-sourcing observations, and constructs related causal graphs and triplets. This paper evaluates multiple open-source LLMs using the collected datasets, benchmarking LLM's ability to predict causal relationships, and to estimate average treatment effect.

Strengths

1. The construction of the causal graph and the estimation of casual relationships are strict and sound. The whole dataset provides high-quality causal relationship annotations for real-life events. 2. A large number of open-source models are evaluated. The evaluation on ATE estimation also involves multiple estimation methods using LLMs.

Weaknesses

1. Human performance is not provided in the empirical comparisons. This makes it hard to understand LLM's performances. One can argue that the performances are far from perfect. However, due to the inherent ambiguity of natural languages, and inherent disagreement in people's opinions, I would assume human performance will also be significantly lower than 100%. Adding human performance in these tables will make the comparisons more informative. 2. Many important details are in the appendix. I understand this is a dense paper with lots of content, however, the presentation and organization can be greatly improved. Additionally, the current results section includes a significant part on how to estimate ATE, which should be included in previous sections. 3. While the dataset itself is huge, there are only five different events (shown in Table 2). Therefore, it is unclear how representative the model's performance on this dataset will be. This is a significant limitation, especially due to the "Causal Parrots" phenomenon mentioned in the introduction.

Questions

1. How sensitive are the eval results w.r.t. the prompts? I'm a bit concerned that since "cause" and "effect" are not common words (especially their formal definition in causal inference), the model's ability may be underestimated with these prompts. 2. There are multiple different methods to estimate probability prediction from LLMs? Have you tried other methods and will that change the empirical results significantly?

Rating

5

Confidence

2

Soundness

3

Presentation

2

Contribution

3

Limitations

Limitations are discussed in Sec. 6.

Replying to the Rebuttal of Authors

Thanks for your detailed response. Some main concerns still remain, so I intend to keep my rating unchanged: * For the causation of necessity, "boarding a plane" is not a cause of "waiting at the luggage belt", while for the causation of sufficiency, they have a causal relationship. So if you do not provide the event "not checking in luggage", you cannot say there is not a causal relationship. When compared to "checking in luggage", "checking in luggage" indeed has a larger probability of becoming a cause of "waiting at the luggage belt". If I provide an event such as "forgetting to collect the luggage", "fight is canceled", or "falling asleep and missed the flight", the causal relationship between "checking in luggage" and "waiting at the luggage belt" does not exist. I insist that causality is about probability if we cannot capture all confounders. * About the human evaluation: I still think human evaluation is important, even language models do not perform well in temporal settings. If humans can easily infer the correct answer by temporal order or necessary relationship between events, then it can be a shortcut to be utilized by language models. By the way, I do not see any human evaluation results in your response to reviewer iBzv.