Rationalization is a self-explaining framework for NLP models. Conventional work typically uses the maximum mutual information (MMI) criterion to find the rationale that is most indicative of the target label. However, this criterion can be influenced by spurious features that correlate with the causal rationale or the target label. Instead of attempting to rectify the issues of the MMI criterion, we propose a novel criterion to uncover the causal rationale, termed the Minimum Conditional Dependence (MCD) criterion, which is grounded on our finding that the non-causal features and the target label are \emph{d-separated} by the causal rationale. By minimizing the dependence between the unselected parts of the input and the target label conditioned on the selected rationale candidate, all the causes of the label are compelled to be selected. In this study, we employ a simple and practical measure of dependence, specifically the KL-divergence, to validate our proposed MCD criterion. Empirically, we demonstrate that MCD improves the F1 score by up to $13.7\%$ compared to previous state-of-the-art MMI-based methods. Our code is available at: \url{https://github.com/jugechengzi/Rationalization-MCD}.
Paper
Similar papers
Peer review
Summary
In this paper, the authors focus on self-explaining rationalization and they propose a Minimum Conditional Dependence (MCD) criterion to select causal rationales from a causal perspective.
Strengths
1. Rationalization is a worthwhile direction to explore in interpretable research. 2. The paper is well written.
Weaknesses
1.This paper lacks novelty. First of all, the causalities in rationalization has been widely discussed[1][2][3]. Besides, the method proposed by the authors is very similar to the previously proposed method[4][5]. Authors do not compare them as baselines and not cite these research paper. 2. The datasets in this paper are limited and I would expect to see more datasets such as ERASER[6] that have been widely used for rationalization[7][8][9]. Furthermore, both BeerAdvocate and HotelReview seem to focus on the binary classification, have the authors studied the effectiveness of MCD on multi-classification tasks, such as DARE[10] and InfoCAL[5] do? 3. Can the authors provide an example to show that the rationales extracted by MCD extraction is not affected by spurious correlations in the data. 4. The recent work is insufficient where much of the recent work on the rationalization is not discussed, such as [3][4][5][6][8][10]. References [1] Invariant rationalization. [2] Interventional rationalization. [3] Discovering Invariant Rationales for Graph Neural Networks. [4] How do Decisions Emerge across Layers in Neural Models? Interpretation with Differentiable Masking [5] Learning from the Best: Rationalizing Prediction by Adversarial Information Calibration [6] ERASER: A benchmark to evaluate rationalized NLP models. [7] An information bottleneck approach for controlling conciseness in rationale extraction. [8] Unifying model explainability and robustness for joint text classification and rationale extraction. [9] UNIREX: A Unified Learning Framework for Language Model Rationale Extraction [10] DARE: Disentanglement-augmented rationale extraction
Questions
1.Why do the authors use different baselines in BeerAdvocate and HotelReview? 2.The authors' discussion of rationalization+LLMs is insufficient, I would like to see more of the authors' views on rationalization+LLMs. 3. The authors' experiments focus on the text classification task, would MCD be equally effective on other nlp tasks, such as text generation?
Rating
3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.
Confidence
5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
Soundness
3 good
Presentation
3 good
Contribution
2 fair
Limitations
See Weaknesses for details.
Summary
This paper studies selective rationalization. Many methods use the maximum mutual information (MMI) criterion to find the most indicative rationale to explain a target label. As it has been shown in the past, this criterion is, by design, sensitive to spurious correlation. This paper proposes a novel criterion instead of "fixing" MMI, called Minimum Conditional Dependence (MCD). The authors identify two stages from which spurious correlations may come from and propose a causal framework to circumvent them. Method The paper assumes that the beer dataset has been generated on 3 aspects: Appearance, Taste, and Smell. This is wrong since the dataset has 5 aspects in total (palate and overall). Nevertheless, the motivating example is convincing. Leveraging the concept of d-separation is novel. However, it is unclear how dissimilar is the proposed approach w.r.t. [1,2] for example (L240 and L130 don't discuss really the differences) A downside of the approach is the need to train one model per aspect. It would interesting to discuss about extending the proposed model. In the experiment section, I would appreciate having the results for the other aspects, and potentially, the original beer dataset to understand the effect of MCD (while not many papers utilize the original beer dataset, I do think it's important). The number of baselines is not consistent through the experiments: T3-T4 would require more models to assess the superiority of MCD. Finally, I am surprised that MCD performs on average better rationalization performance but worst predictive performance, which defeats the purpose of "rationalization": explaining the output of a model. Why is there such a gap? The related work is quite complete. Nevertheless, I have identified few missing citations [3-5]. 1 Chang et al. 2021, Invariant Rationalization (ICML) 2 Yue et al. 2022, Interventional Rationalization 2 Chang et al. 2019, A Game Theoretic Approach to Class-wise Selective Rationalization (NeurIPS) 3 Antognini et al. 2021, Multi-Dimensional Explanation of Target Variables from Documents (AAAI) 4 Antognini and Faltings 2021, Rationalization through Concepts (ACL)
Strengths
- The causal framework is novel, and using d-separation is a nice replacement to circumvent MMI's flaws - MCD obtains better rationalization performance - The paper is well written and motivated.
Weaknesses
- It is unclear how dissimilar is the proposed approach w.r.t. [1,2] for example (L240 and L130 don't discuss really the differences) - The proposed model only works for 1 aspect. - MCD obtains worst predictive performance, which defeats the purpose of "rationalization": explaining the output of a model.
Questions
- Why MCD underperforms in terms of predictive performance? - How could we extend MCD to multi-dimensional rationales? - Could you report results of MCD with BERT for the other experiments since it provides better performance?
Rating
5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.
Confidence
4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
Soundness
2 fair
Presentation
3 good
Contribution
2 fair
Limitations
The authors do not discuss about the limitations and potential negative societal impact of their work.
Summary
The maximum mutual information criterion is commonly used for rationalization, but it uncovers associations rather than causal relationships. The authors propose to identify non-causal features that are independent of the labels given the causal features and the ‘minimum conditional dependence’ criterion, which does not require prior expert knowledge. Experiments are focused on the popular BeerAdvocate and HotelReviews datasets, though mostly on the former, and obtain very competitive results against the other, well-chosen baselines.
Strengths
- The treatment of this problem using causal terminology is appropriate and interesting, and attempts were made to formalize some of these concepts (in Sec 4.2), with only some challenges, listed below. - Appropriate baselines and datasets are provided in the experiments, as per other related (uncited) work. Moreover, attempting to overcome limitations of ‘some of the baseline methods’ (L290) by establishing consistent ‘settings’ is good, although what those settings are (beyond using GloVE) should be included. - The rates of improvement in Table 2 are impressive. In some cases, tests of statistical significance would be useful to include, at least against FR, though apparently not necessary.
Weaknesses
- The treatment of related works is extremely superficial in Section 2. Approximately half of the section on rationalization is merely a list of papers barely described by their topics and summarized in whole as being ‘orthogonal’. To some extent, a literature survey was better covered in Section 1, - There are various minor issues including: - atypical English (L2 “pieces of their inputting texts”; L48 “is easily to be affected”; L56 “[lowercase] comments regarding”; L188 “a image”) - very small text (Fig 1) - incomplete citations (e.g., Yue et al (2023), L514.)
Questions
- This seems like a companion piece to “MGR: Multi-generator Based Rationalization” by Liu et al (2023), as much of the text is identical and they also deal with spurious correlation and degeneration, but the focus is on the MCD criterion rather than using multiple generators. What other differences can easily be described in your paper, if this paper were cited? - Is your ‘derivation’ in appendix B1 really an example of Bayes’ theorem, or just an application of marginalization and the chain rule? - Assumption 1 (L223) should be better explained in Appendix B, as the ‘temporal sequence’ explanation provided in the paper is insufficient, especially if it refers to some aspect of dataset acquisition or prediction rather than the actual state of the world. Can you expand on this? - the FR method (L274) is claimed to be the SotA, but even the paper by Liu et al (2022) to which it refers does not make a compelling case for it being the true SotA — how is this claim proven conclusively? - Given the new methodological approach, it would be interesting to include some study of the required computational resources in the main body of the text, beyond the brief mention of RTX3090 and appendix A.5’s epochs, including possible ablations. Would that be possible? - Can you add an explanation (or rationale) for why fine-tuning BERT would be ‘challenging’ (L299) beyond just providing citations and some indication of potential overfitting (section beginning L331)?
Rating
7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.
Confidence
4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
Soundness
3 good
Presentation
3 good
Contribution
3 good
Limitations
- The last paragraph of Sec 6 touches briefly on the limitations. It is suggested that not computing the ‘precise values’ of causal effects would be a limitation, although it is not clear why that would be. Various other forms of limitation, including the relatively restricted datasets (in terms of task, scope, and number) used in the experiments or the focus on text, could also be addressed, for example.
Summary
This work focuses on the problem of rationalization that aims to extract explanatory sentences that serve as rationales along with training a predictor for the downstream task. Prior work on rationalization has typically employed the maximum mutual information (MMI) criterion. However, this criterion does not uncover causal relationships and latches on to feature correlations and degenerations. The key insight in this work is that non-causal features are independent of the target labels given the causal features, based on which the authors propose a minimum conditional dependence criterion to discover causal rationales.
Strengths
- Paper is written well. - Offer a unified perspective of feature correlations and degenerations (described in prior work on rationalization). - Propose a new MCD criterion to identify causal rationales. - Outperform existing MMI-based rationalization techniques on two multi-aspect sentiment classification tasks.
Weaknesses
One of the main contributions of this work is that the proposed MCD criterion leads to the discovery of causal rationales. It would be useful for the reader to see some examples of generated rationales via MCD, and also highlight failure cases when MCD fails to retrieve the causal rationales.
Questions
- Why is Equation 16 a good approximation? Define \Omega(M) in Equation 16. - In Table 5, why are the F1 scores of FR-ELECTRA so poor on the Appearance and Palate aspects? - Some anecdotal examples of causal rationales that were identified via MCD as opposed to MMI would be useful for the reader.
Rating
7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.
Confidence
3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
Soundness
3 good
Presentation
3 good
Contribution
3 good
Limitations
Some limitations have been listed.
Further clarifications on the novelty and datasets.
**A1:** Dear reviewer, I'd like to further elaborate on the distinct novelty of our work in comparison to [4]. Before we begin, we'd like to emphasize that our primary contribution lies in the novel theoretical insights regarding rationale causality and the introduction of the MCD criterion as a replacement for MMI, rather than in a specific model (L105-109). Below are some distinctions from the perspective of the proposed model: While [4] focuses on post-hoc explanation, our study concentrates on self-rationalization. These are two distinct research areas. The primary challenge in post-hoc explanation is ensuring faithfulness (see L27-33 of our submission and Section 3 of [4]), which means the explanation should correspond closely with the model's prediction, $\hat{Y}$. This intuitively leads to aligning $P(\hat{Y}|X)$ with $P(\hat{Y}|X_Z)$. But in our paper, we are extracting rationales that are causal to the human-annotated **gold label $Y$**. Note that from a theoretical standpoint, even subtle differences can lead to very different results. Very different from [4], the goal of our Equation 16 is designed to align $P({Y}|X)$ and $P({Y}|X_Z)$ (L252). This necessitates minimizing the cross-entropies to approximate both $P({Y}|X)$ and $P({Y}|X_Z)$ via $P(\hat{Y}|X)$ and $P(\hat{Y}|X_Z)$, respectively (L253-257). The above motivations are all based on our causality analysis, which is central to our research. [4] doesn't touch on causality analysis, and as a result there is no approximation for $P({Y}|X_Z)$ in [4]. Here are the ablation experiments that remove the approximation for $P({Y}|X_Z)$: | Appearance | S | Acc | P | R | F1 || S | Acc | P | R | F1 ||S | Acc | P | R | F1 | |:---:|---|---|---|---|---|-|---|---|---|---|---|-|--|---|---|---|---| | MCD | 9.5 | 81.5 | 94.2 | 48.4 | **63.9** || 20.0 | 85.5 | 79.3 | 85.5 | **82.3** || 29.7| 86.7 | 59.6 | 95.6 | **73.4** | | w/o $H_c(Y,\hat{Y}_z\|X_Z)$ | 9.7 | 72.5 | 85.1 | 44.6 |58.5 || 20.0 | 75.5 | 74.6 | 80.7 |77.5 || 32.4 | 81.2 | 54.6 | 95.5 |69.5 | | Aroma | S | Acc | P | R | F1 || S | Acc | P | R | F1 || S | Acc | P | R | F1 | |:---:|---|---|---|---|---|-|---|---|---|---|---|-|---|---|---|---|---| | MCD | 9.9 | 87.5 | 84.6 | 53.9 | **65.8** || 19.3 | 88.4 | 65.8 | 81.4 | **72.8** || 29.6 | 90.2 | 46.1 | 87.5 | **60.4** | | w/o $H_c(Y,\hat{Y}_z\|X_Z)$ | 11.4 | 86.3 | 70.2 | 51.6 |59.4 || 21.4 | 84.3 | 58.8 | 80.6 |68.0 || 30.4 | 88.4 | 40.4 | 78.8 |53.4 | | Palate | S | Acc | P | R | F1 || S | Acc | P | R | F1 || S | Acc | P | R | F1 | |:---:|---|---|---|---|---|-|---|---|---|---|---|-|---|---|---|---|---| | MCD | 9.4 | 87.3 | 60.9 | 47.1 | **53.1** || 19.6 | 87.7 | 41.3 | 65.0 | **50.5** || 29.4 | 87.0 | 30.5 | 72.4 | **42.9** | | w/o $H_c(Y,\hat{Y}_z\|X_Z)$ | 10.6 | 83.7 | 53.0 | 45.2 |48.8 || 20.5 | 85.2 | 37.3 | 61.4 |46.4 || 31.5 | 86.5 |25.4 | 64.4 |36.4 | We see that when we remove the approximation, there is a significant drop in the F1-score, demonstrating the importance of our novel theoretical analyses. We are grateful for your time and expertise in reviewing our paper. To better address your feedback, could you elaborate on the similarity concerns you raised? **A2:** Here are some further clarifications about the ERASER datasets. We think there is a misunderstanding. While ERASER is indeed a valuable resource, it may not always be the best benchmark for every research problem. The primary strength of ERASER datasets is the inclusion of human-annotated rationales in the training set, making them particularly beneficial for supervised rationale extraction. To the best of our knowledge, the majority of methods that leverage ERASER datasets, including those you've mentioned ([7,8,9]), all require human-annotated rationales for training. In contrast, methods that focus on unsupervised rationale extraction, such as [1,2,5,10] that you mention, tend not to employ ERASER. Currently, the Beer dataset remains a predominant choice in this area. We opted for the Beer and Hotel datasets, aligning with the recent and strong baseline FR from NeurIPS 2022.
Response to rebuttal
Thanks to the authors for their detailed response (including new results) and clarifications to my questions. I'm raising my score to 7.
Thank you very much for taking the time to review our paper! With best wishes to you and yours!
Response to rebuttal
Thanks for the response. But I still have the following questions: **about A2** I do not agree with the authors' explanation for not using the ERASER dataset. Since MCD should eventually be applied in practice, it should be validated on more datasets to verify the effectiveness of MCD. Moreover, the datasets in ERASER all contain true rationale labels and can be considered as the causal rationale. **about A4** I don't agree with the authors' statement "Given the vastness of the rationalization field, it's hard to cover all new papers in a non-survey paper". Some of the relevant work I mentioned in the review comments was done before 2022. For NeurIPS 2023, they are not new papers.
We are grateful for your feedback.
We are grateful for your feedback. Here are some further clarifications. **A1 (further clarification on dataset selection):** We agree that the application is important. But as a research paper, the most important role of the experiments is to verify the theoretical claims, rather than to achieve the engineering SOTA. This is because different applications face different challenges, and different methods are aimed at solving different problems. We can choose different methods for different applications. As implied in the previous rebuttal, the most appropriate dataset for verifying the ability to select causal rationales is still the BeerAdvocate, and the methods that are most relevant to our research (INVRAT, Inter_RAT, and FR) all use BeerAdvocate as their main experiment. Also, the datasets we chose are just the same as the strongest baseline FR (NeurIPS 2022). There are two lines of research in rationalization: supervised rationalization and unsupervised rationalization, where "supervised" means that human-annotated rationales are required for training. As implied in the previous rebuttal, the primary advantage of the ERASER datasets is the inclusion of human-annotated rationales in the training set, making them particularly useful for supervised rationale extraction. To the best of our knowledge, the majority of methods that leverage the ERASER datasets, including those you've mentioned ([7,8,9]), are all supervised methods. However, our research is unsupervised rationalization, and the datasets we chose are those widely used by other unsupervised methods. Our datasets align with two recent papers DMR (AAAI2021) and FR (NeurIPS 2022). And we recently found that a recent ( June 25, 2023) published paper CR[A] also uses BeerAdvocate as the main experiment and also uses HotelReviews as a supplement, without using ERASER, which invalidates the usefulness of the BeerAdvocate and HotelReviews datasets we used. **A2 (new references):** Thank you for your suggestion. We agree that some of the papers in the field of rationalization were not covered in the original submission, and we will discuss them in Sec.2 in the next version. What I mean by "it's hard to cover all new papers in a non-survey paper" is that we only discussed the papers that study similar research topics (i.e., feature correlation and degeneration) with us in detail, and ignored those sub-relevant papers, due to the page limit. We appreciate these valuable references and we will follow up on your suggestion to make a more comprehensive survey. We appreciate your continued dedication and effort in evaluating our manuscript. [A] Towards Trustworthy Explanation: On Causal Rationalization. ICML 2023.
Thank you for the author's response.
Thank you for the author's response. However, the author's reply has not addressed my concerns, including the novelty and comprehensiveness of the experiments. For the ERASER dataset, both the test set and training set have manually annotated rationales, while the beer dataset only has manually annotated rationales in the test set. I acknowledge the usefulness of the beer dataset, but I still insist that the author should conduct experiments on the ERASER benchmark dataset. Furthermore, the author's understanding of [7] is incorrect. IB is still an unsupervised rationalization, and it is highly scalable, capable of extending to supervised rationalization. However, its core part (Figure 2 in [7]) is an unsupervised rationalization. I hope the author can carefully read the work I mentioned and improve the related work section.
Thank you for the timely feedback. And could you please provide more details about the novelty concerns?
We appreciate your suggestions and are already conducting experiments on one of the ERASER datasets (i.e., Movie), which will be completed soon. Could you elaborate more on your concerns about novelty? If you could pinpoint the specific similarities between our paper and others, we believe we can address your questions more effectively. We are aware that the framework in [7] can be adapted for both supervised and unsupervised settings. However, in the experiments of [7], they employed a semi-supervised setup where 25% of the training data requires human-annotated rationales. This necessitates the use of the ERASER dataset, which is why we categorize it under supervised methods. We appreciate you pointing this out and will read the new references carefully to improve the related work section.
Additional experiments on ERASER.
We have now performed additional experiments on one of the ERASER datasets: **Movie**. And here are the results: | Method | P | R | F1 | |:---:|---|---|---| | Inter_RAT\* | 35.7 | 35.8 | 35.7 | | FR | 36.0 | 36.4 | 36.2 | | MCD (Ours) | 37.6 | 38.9 | **38.2** | Note that INVRAT relies on the labels of other aspects to create different environments to remove feature correlations, so we cannot compare with INVRAT on this non-multi-aspect classification dataset. However, we can see that INVRAT performs worse than Inter_RAT and FR on the Beer dataset. So on this dataset, we only compare with Inter_RAT and FR. "\*": The results of Inter_RAT are obtained from its original paper. We follow Inter_RAT to set the rationale sparsity to be about 20%. Since there are not many feature correlations in this dataset, we don't get as much improvement as in the Beer and Hotel datasets. But mask correlations can still occur in this dataset, so our MCD still gets considerable improvements over MMI-based methods ($\frac{38.2-36.2}{36.2}\approx$5.5%). Here is a list of the papers that are close to our research topics (i.e., feature correlation or degeneration) and their corresponding datasets: | Method | Venue | Datasets | |:---:|---|---| | INVRAT | ICML 2020 | Beer+IMDB | | Inter_RAT | unpublished| Beer+Movie+LJP | | DMR | AAAI 2021 | Beer+Hotel | | A2R | NeurIPS 2021 | Beer+Movie | | FR | NeurIPS 2022 | Beer+Hotel | | CR [A] | ICML 2023 | Beer+Hotel+GA | | MGR [B] | ACL 2023 | Beer+Hotel | | DR [C] | SIGKDD 2023 | Beer+Hotel | Notes: LJP and GA datasets have no human-annotated rationales for testing. Our choice of datasets (Beer and Hotel) seems reasonable, as this choice is also widely used by other methods. Thank you very much for taking the time to review our paper! With best wishes to you and yours! [A] Towards Trustworthy Explanation: On Causal Rationalization. ICML 2023. [B] MGR: Multi-generator Based Rationalization. ACL 2023. [C] Decoupled Rationalization with Asymmetric Learning Rates: A Flexible Lipschitz Restraint. SIGKDD 2023.
Thank you very much for taking the time to review our paper! With best wishes to you and yours!
Thank you very much for taking the time to review our paper! With best wishes to you and yours!
Decision
Accept (poster)