Is the MMI Criterion Necessary for Interpretability?

Degenerating Non-causal Features to Plain Noise for

Self-Rationalization

Wei Liu1 Zhiying Deng1∗ Zhongyu Niu1 Jun Wang2∗

Haozhao Wang1∗ YuanKai Zhang1 Ruixuan Li1∗
arXiv:2410.06003v4 [cs.LG] 22 Oct 2024

1
School of Computer Science and Technology,
Huazhong University of Science and Technology
2
iWudao Tech
1
{idc_lw, dengzhiyingdd, zy_niu, hz_wang, yuankai_zhang, rxli}@hust.edu.cn
2
[email protected]

Abstract

An important line of research in the field of explainability is to extract a small

subset of crucial rationales from the full input. The most widely used criterion for
rationale extraction is the maximum mutual information (MMI) criterion. However,
in certain datasets, there are spurious features non-causally correlated with the
label and also get high mutual information, complicating the loss landscape of
MMI. Although some penalty-based methods have been developed to penalize the
spurious features (e.g., invariance penalty, intervention penalty, etc) to help MMI
work better, these are merely remedial measures. In the optimization objectives of
these methods, spurious features are still distinguished from plain noise, which hin-
ders the discovery of causal rationales. This paper aims to develop a new criterion
that treats spurious features as plain noise, allowing the model to work on datasets
rich in spurious features as if it were working on clean datasets, thereby making ra-
tionale extraction easier. We theoretically observe that removing either plain noise
or spurious features from the input does not alter the conditional distribution of the
remaining components relative to the task label. However, significant changes in
the conditional distribution occur only when causal features are eliminated. Based
on this discovery, the paper proposes a criterion for Maximizing the Remaining
Discrepancy (MRD). Experiments on six widely used datasets show that our MRD
criterion improves rationale quality (measured by the overlap with human-annotated
rationales) by up to 10.4% as compared to several recent competitive MMI variants.
Code: https://github.com/jugechengzi/Rationalization-MRD.

1 Introduction
With the success of deep learning, there are growing concerns over interpretability (Lipton, 2018).
Ideally, the explanation should be both faithful (reflecting the model’s actual behavior) and plausible
(aligning with human understanding) (Jacovi and Goldberg, 2020; Chan et al., 2022). Post-hoc
explanations, which are trained separately from the prediction process, may not faithfully represent an
agent’s decision, despite appearing plausible (Lipton, 2018). In contrast to post-hoc methods, ante-hoc
∗
Corresponding authors.
This paper is a collaboration between Intelligent and Distributed Computing Laboratory, Huazhong University
of Science and Technology and iWudao Tech. The companion pieces to this research series can be found at
https://jugechengzi.github.io/WeiLiu.github.io//English/.

38th Conference on Neural Information Processing Systems (NeurIPS 2024).

(or self-explaining) techniques typically offer increased transparency (Lipton, 2018) and faithfulness
(Yu et al., 2021), as the prediction is made based on the explanation itself. There is a stream of
research that has exposed the unreliability of post-hoc explanations and called for self-explanatory
methods (Rudin, 2019; Adebayo et al., 2018; Ghassemi et al., 2021; Ren et al., 2024).
One important line of research to build self-explainable NLP models is first extracting the most
informative rationale in a text and then using the extracted rationale to train a predictor. This
line of research is known as rationalization. A model-agnostic rationalization framework, called
Rationalizing Neural Predictions (RNP), was first proposed by Lei et al. (2016). RNP utilizes a
cooperative game between an extractor and a predictor. This game is designed with a focus on
"data-centric" importance of rationales (i.e., it aims to explain the connection between a text and the
model-agnostic task label, rather than explaining the output of a specific model). First, the extractor
identifies the most informative part of the input, known as the rationale. Then, as depicted in Figure 1,
the rationale is transmitted to the predictor to facilitate predictions. The extractor and predictor
are trained cooperatively to maximize prediction accuracy, with the theoretical support being the
Maximum Mutual Information (MMI) criterion (Yu et al., 2021; Chang et al., 2020). RNP and its
variants have become mainstream approaches for enhancing the interpretability of NLP models (Liu
et al., 2022, 2023c,a; Storek et al., 2023; Zhang et al., 2023; Liu et al., 2024; Zhao et al., 2024; Jiang
et al., 2024; Hu and Yu, 2024; Yue et al., 2024). Aside from interpretability, rationalization can also
serve as a method for data cleaning, as the extracted (Z, Y ) samples can function as a new dataset.
Recent studies have shown that a predictor trained with such a dataset can be more robust (Chen et al.,
2022) and generalizable (Wu et al., 2022; Gui et al., 2023), due to the removal of task-irrelevant,
harmful information.
Previous methods typically
employ the Maximum Mu- 𝑌 [0, 1]
tual Information (MMI) cri-
Extractor

Predictor
great hotel with great hotel with
excellent location ! excellent location !
0.1
terion to identify the ratio- the breakfast buffet the breakfast buffet ෠
𝑚𝑖𝑛θ𝑃,θ𝐸 𝐻𝑐 (𝑌, 𝑌)
nale, defined as the subset is terrific . service is terrific . service
0.9
is wonderful . is wonderful .
most indicative of the tar-
get label. However, cer- 𝑿 𝑍 𝑌෠
tain datasets contain fea-
tures that are statistically Figure 1: The standard rationalization framework RNP. The task is
correlated with the task la- binary sentiment classification about the hotel’s location. X, Z, Ŷ , Y
bel but do not causally af- represent the input, the extracted rationale candidate, the prediction and
fect it. These features are the ground truth label, respectively. θE , θP represent the parameters of
referred to as spurious fea- the extractor and the predictor, respectively. Hc denotes cross-entropy.
tures, and the associated
correlations are known as spurious correlations. The spurious features are also indicative of the target
label and can compete with the true rationale for extraction opportunities under the MMI criterion,
distinguishing them from plain noise. Consider a scenario where the extractor is initially positioned
on selecting plain noise. If a clean dataset contains no spurious features, the gradient will guide
the extractor solely towards causal features. However, if the dataset is rich in spurious features, the
extractor can move in various directions, arbitrarily towards either spurious or causal features. Given
the potential diversity of spurious features in the data, the extractor may struggle in a complex loss
landscape (Chang et al., 2020). A typical example of spurious correlation, as highlighted in LIME
(Ribeiro et al., 2016), is the frequent co-occurrence of wolves and snow in images. Consequently,
the presence of snow in the background can erroneously serve as a strong indicator for classifying
an image as depicting a wolf, leading MMI to possibly select the background feature instead of the
wolf’s face as the rationale. Figure 5(a) in Appendix A.10 illustrates another instance of spurious
correlations.
Some methods try to develop regularizers that can penalize the spurious features and fix the shortcom-
ing of MMI. INVRAT (Chang et al., 2020) incorporates the concept of invariant risk minimization
to design an invariance penalty. Inter_RAT (Yue et al., 2023) utilizes an intervention penalty. CR
(Zhang et al., 2023) implements a sufficiency and necessity penalty by separately assessing the
sufficiency and necessity of each token. In addition to the specific shortcomings of each type of
method (discussed in §2), they share a common limitation: most still adhere to the MMI criterion and
merely use supplementary objectives to penalize spurious features. If the penalty term’s weight is
too small, spurious features will still be favored over uninformative noise due to their higher mutual

2
information. Consequently, when the extractor initially selects noise, the gradient descent algorithm
might shift towards either spurious features or the true rationale. On the other hand, if the penalty
term’s weight is too high, it can dominate the loss function and impair the MMI’s ability to distinguish
between noise and causal features (see §4.1). The difference between spurious features and noise
can complicate the loss landscape of rationale extraction, which may lead to the emergence of local
optima. Note that the problem of local optima in rationalization is very serious (Yu et al., 2021).
A recent research MCD (Liu et al., 2023b) revises the MMI criterion to the minimum conditional
dependence criterion and does not introduce extra penalty terms. However, MCD also can not promise
to treat the spurious features as plain noise. We provide a detailed comparison of MCD and our
approach in Appendix A.1 help readers better understand the unique advantages of our approach.
In this paper, we diverge from previous research that focuses on the selected rationale candidate
Z as the primary subject. Instead, we adopt a reversed perspective, considering the remaining
part X−Z by excluding the rationale candidate Z from the full input X, as the main subject of
study. We find that, although selecting spurious features rather than noise as Z will be more
indicative of Y (i.e., P (Y ∣S) ≠ P (Y ∣N ), with S, N denoting Spurious features and Noise), neither
selecting the plain noise nor the spurious features as Z will cause a change in P (Y ∣X−Z ) (i.e.,
P (Y ∣X−S ) = P (Y ∣X−N ) = P (Y ∣X)). Based on this observation, we replace the criterion of
maximizing the mutual information I(Y ; Z) with maximizing the remaining discrepancy (MRD)
DKL (PY ∣X ∣∣PY ∣X−Z ). Under this new criterion, spurious correlations are treated as equivalent to
uninformative noise without extra supplement regularizers on the rationale candidate, allowing the
extractor to work on datasets rich in spurious features as if it were working on clean datasets.
In summary, our contributions are as follows: (1) We introduce a new criterion that treats spurious
features as equivalent to plain noise, simplifying the loss landscape for rationale extraction. (2) We
propose a simple and practical method to implement this new criterion. (3) Experiments on six
widely used datasets show that our MRD improves the rationale quality (measured by the overlap
with human-annotated rationales) by up to 10.4% as compared to several competitive MMI variants.

2 Related work

Data-centric rationale extraction. Data-centric rationale extraction (also known as rationalization)

is a general framework first proposed by Lei et al. (2016). By extracting rationales before making
predictions, this framework has been one of the mainstreams to facilitate the interpretability of NLP
models (Chang et al., 2020; Sha et al., 2021; Yu et al., 2021; Shen et al., 2022; Chan et al., 2022;
Storek et al., 2023; Zhang et al., 2023). And DeYoung et al. (2020) proposed a benchmark that can
be used for supervised rationale extraction. Recently, there has also been some work attempting to
extend it to the field of graph learning (Luo et al., 2020) and computer vision (Yuan et al., 2022).
Apart from improving interpretability, recent work has also discovered that it can serve as a method
of data cleaning, as training a predictor with the extracted rationales has been found to increase
robustness (Chen et al., 2022) and generalization (Wu et al., 2022; Gui et al., 2023). We also briefly
discuss the potential impact of rationalization in the era of LLMs in Appendix A.11.
Mitigating spurious correlations. One important obstacle of rationalization is the spurious features
in datasets, as the spurious features also have high correlations with the task label and can compete
with the causal features for extraction opportunities under the most widely used MMI criterion. Some
methods have been developed to mitigate the impact of spurious correlations. INVRAT (Chang et al.,
2020) attempts to tackle feature correlation using invariant risk minimization (IRM) (Arjovsky et al.,
2019). The main idea is to penalize spurious (non-causal) variations by splitting the dataset into
distinct environments. However, IRM-based methods have several limitations. For instance, they
require strong prior knowledge about the relationships between non-causal and causal features (e.g.,
the extra labels of non-causal features) in order to divide the dataset (Lin et al., 2022b). Moreover,
IRM-based methods are limited to addressing only a finite set of predetermined non-causal features,
neglecting the potential existence of numerous unknown non-causal features. In fact, a recent study
(Lin et al., 2022b) in the field of IRM has theoretically demonstrated that it is nearly impossible
to partition a dataset into different environments to eliminate all non-causal features using IRM.
Other challenges, such as the tendency to overfit, difficulty in applying to larger models (Zhou et al.,
2022; Lin et al., 2022a), and the marginal shift risk of the input (Rosenfeld et al., 2021), have also
been identified within the realm of IRM. Inter_RAT (Yue et al., 2023) attempts to eliminate feature
correlation through backdoor adjustment, intervening directly with the confounders. However, it is

3
extremely hard to measure the confounders since they are usually not observable in the dataset. CR
(Zhang et al., 2023) calculates the sufficiency and necessity of each token separately, which leads to a
high computational complexity, making it feasible only for very short texts. Aside from the above
shortcomings, penalty-based methods share a common limitation. They need to coordinate the MMI
and the penalty objectives to make the gradient descent algorithm treat the spurious features and plain
noise equally and guide the extractor to move towards only the causal features. A recent research
MCD (Liu et al., 2023b) revises MMI to the minimum conditional dependence criterion. Although
MCD does not involve penalty regularizers, it also cannot treat spurious features and plain noise
equally. And the spurious features can still compete with the causal ones.

3 Preliminaries
3.1 The rationale extraction task

We consider the text classification task, where the input is a text sequence X=[x1 , x2 , ⋯, xl ] with
xi being the i-th token and l being the number of tokens. Y represents the classes in a dataset D.
The standard rationalization framework RNP (Lei et al., 2016) consists of an extractor fE (⋅) and
a predictor fP (⋅), with θe and θp representing the parameters of the extractor and predictor. For
(X, Y ) ∼ D, the extractor first outputs a sequence of binary mask M = fE (X) = [m1 , ⋯, ml ] ∈
{0, 1}l (in practice, the extractor first outputs a Bernoulli distribution for each token and the mask for
each token is independently sampled using gumbel-softmax). Then, it forms the rationale candidate
Z by the element-wise product of X and M :
Z = M ⊙ X = [m1 x1 , ⋯, ml xl ]. (1)
To simplify the notation, we denote fE (X) as Z in the following sections, i.e., fE (X) = Z. With the
extractor’s selection, we get a set of (Z, Y ) samples, which are generally considered to represent the
distribution P (Y ∣Z). The rationale Z is searched by maximizing the mutual information I(Y ; Z):

Z ∗ = arg max I(Y ; Z) = arg max(H(Y ) − H(Y ∣Z)) = arg min H(Y ∣Z), s.t., Z = fE (X). (2)
Z Z Z

In practice, the entropy H(Y ∣Z) is commonly approximated by the minimum cross-entropy
minθp Hc (Y, Ŷ ∣Z), with Ŷ = fP (Z) representing the output of the predictor. It is essential to
note that the minimum cross-entropy is equal to the entropy (please refer to Appendix A.7). Replac-
ing Z with fE (X), the extractor and the predictor are trained cooperatively:
min Hc (Y, fP (fE (X))∣fE (X)), s.t., (X, Y ) ∼ D. (3)
θe ,θp

To make the selected rationale human-intelligible, rationalization methods usually constrain the
rationales by compact and coherent regularization terms. In this paper, we use the most widely used
constraints proposed by Chang et al. (2020):
l
∣∣M ∣∣1
Ω(M ) = λ1 ∣ − s∣ + λ2 ∑ ∣mt − mt−1 ∣. (4)
l t=2
The first term encourages that the percentage of the tokens being selected as rationales is close to a
pre-defined level s. The second term encourages the rationales to be coherent.

3.2 Causality

We note that the contribution of this part does not belong to this paper. To help readers unfamiliar
with causality better understand the spurious correlations, we borrow it from a previous paper MCD
(Liu et al., 2023b) and make some minor revisions to make this paper self-contained. We provide a
detailed comparison with MCD in Appendix A.1.
We consider that X consists of a set of variables {N, S, C}, where C denotes the real causal rationale
for the corresponding task label Y . And N, S represent the plain Noise and Spurious features,
respectively. The extractor selects one of {N, S, C} to be the rationale candidate Z. Note that Z is
not a separate variable, but a proxy for any variable within X. Initially, the extractor may randomly
select either N, S or C to be Z.

4
 𝐿 𝐿 𝐿 𝐿
𝐿(𝑆)
𝑔(𝑆 → 𝐶)
𝑔(𝑆 → 𝑁)
𝐿(𝑁) 𝐿(𝑁) 𝐿(𝑁) 𝐿(𝑁)
𝑔(𝑁 → S) 𝑔(𝑁 → 𝐶) 𝑔(𝑁 → S) 𝑔(𝑁 → 𝐶) 𝑔(𝑁 → 𝐶)
𝐿(𝑆) 𝐿(𝑆)
𝐿(𝑆)
𝐿(𝐶) 𝐿(𝐶) 𝐿(𝐶) 𝐿(𝐶)

θ𝑆 θ𝑁 θ𝐶 θ𝑆 θ𝑁 θ𝐶 θ𝑆 θ𝑁 θ𝐶 θ𝑁 θ𝑆 θ𝐶
(a) under-penalty (b) no penalty (c) appropriate penalty (d) over-penalty

Figure 3: Penalizing spurious features for more efficiently searching causal rationales.

Consider a classification dataset, we posit a prob-

abilistic graphical model to illustrate the corre- U U
sponding data-generating process in Figure 2(a).
The annotators assign the task label Y by view- S N C XT N XA
ing the causal features in X (C Ð → Y ). There
are also some spurious features non-causally
associated with Y through some unobservable Y Y
confounders U (S ← ÐU Ð →CÐ → Y ). (a) (b)

To facilitate understanding, let’s take a widely Figure 2: The data-generating process of (a) a gen-
used dataset, Beer-Appearance, as an example eral classification dataset and (b) a specific dataset
for a detailed analysis in Figure 2(b). The task Beer-Appearance.
is binary sentiment classification for beer’s ap-
pearance. The input X comprises comments on two aspects (we omit other aspects for brevity):
XT for Taste and XA for Appearance, each of which can be considered as a subset variables of
X. Additionally, N signifies something that does not discuss the sentiment tendency of X. The
annotators assign the appearance label Y by viewing the comments on appearance (XA Ð → Y ).
Therefore, only XA serves as the direct cause for Y . However, XA is correlated with XT due to a set
of unobserved variables U (called confounders). For example, U may include a variable indicating
whether the beer originates from a reputable brand, and a pleasant taste may imply that the beer
comes from a good brand (U Ð → XT ). Moreover, a beer from a reputable brand is likely to have a
pleasant appearance (U Ð → XA ). Consequently, XT is associated with Y via a backdoor path, as
depicted by the red dotted line in Figure 2(b). In this situation, XT is somewhat indicative of Y
(please refer to Appendix A.2 for a quantitative example), but it signifies a statistical correlation
rather than causality. With the objective of MMI (Equation 3), XT can compete with XA for the
opportunity to be selected as the rationale candidate, complicating the rationale extractor’s search
landscape.

4 Treating spurious features as equivalent to plain noise

4.1 The shortcomings of penalty-based MMI

Since spurious features also have a high correlation with the task label, some methods tend to penalize
spurious features with some supplementary regularizers (discussed in §2). Generally, their loss
functions can be written in a form like
L(Z) = LM M I (Z) + λLpenalty (Z), (5)
where Z is the rationale candidate, which is a proxy of the variables within X (e.g., C, S or N ).
We now present some qualitative analysis to demonstrate why using penalties to amend the MMI
criterion can only partially mitigate the issue of spurious correlations. Generally, for the MMI loss,
we have LM M I (C) ≤ LM M I (S) < LM M I (N ) in real-world datasets (please refer to Appendix
A.3 for detailed discussion). For the penalty loss, we usually have Lpenalty (C) < Lpenalty (S) and
Lpenalty (N ) < Lpenalty (S). We denote d(⋅, ⋅) as the distance of the extractor’s parameters moving
from one state to another. For example, d(N, C) denotes the distance between the extractor’s two
states selecting N and C respectively. We denote g(N Ð → C) = L(N )−L(C)
d(N,C)
as the (qualitative)
tendency of the extractor’s moving from N towards C.

5
If λ = 0 (vanilla MMI), although we have g(N Ð → C) > 0, we also g(N Ð → S) > 0. Thus the extractor
may move towards either C or S with gradient descent, not necessarily C (like the situation shown in
Figure 3(b)). This could lead to longer optimization paths, and the additional paths might introduce
extra local optima. Note that Yu et al. (2021) have shown that local optima are serious in unsupervised
(with no human-annotated rationales for supervision) rationalization.
MMI allows the extractor to move towards either spurious features or causal features when starting
from plain noise (Figure 3(b)). Conversely, penalties enable the extractor to move towards either plain
noise or causal features when starting from spurious features (Figure 3(d)). If these two objectives are
well-coordinated such that L(S) = L(N ) > L(C), the loss landscape will be much simpler and the
extractor can ultimately move towards causal features (Figure 3(c)). However, such a coordination
is not each to achieve. If λ is too small, the situation will be under-penalty (Figure 3(a)) and the
spurious features can still compete with the causal features for extraction opportunities. If λ is
too high, the situation can become one of over-penalization (Figure 3(d)), where the influence of
MMI in distinguishing between noise and causal features may be decreased by the domination of
λLpenalty (Z). As a result, noise can compete with causal features for the chance of being selected.
In conclusion, a good objective should make that g(N Ð → C) > 0, g(S Ð → C) > 0, L(S) = L(N ).
Since none of the existing MMI variants can treat spurious features as equivalent to plain noise. It
then leads to a question: is MMI really necessary for rationale extraction? Can we no more use
auxiliary regularizers to fix it, but just remove it completely and replace it with other criteria?

4.2 Spurious features are equivalent to plain noise in a counterfactual view

We aim to develop a new criterion that can treat spurious features as equal to plain noise, so that
regardless of whether the extractor currently selects S or N , the gradient descent algorithm can guide
the extractor to move only towards C. In this paper, we adopt a perspective that reverses common
methods. We no longer focus on the selected rationale candidate as previous methods do. Instead, we
look into the properties of the remaining part after excluding the rationale candidate.
We denote the non-causal subset of X as A = {S, N }. From the probabilistic graphical model shown
in Figure 2(a), we know that A and Y are d-separated by the causal features C (Liu et al., 2023b)
(please refer to Appendix A.6 for a detailed illustration). It means that all variables within A are
independent with Y when conditioned on C.
With the d-separation property, we have P (Y ∣C, S) = P (Y ∣C) = P (Y ∣C, N ) = P (Y ∣C, N, S).
This inspires us to view the problem from a perspective opposite to previous studies; that is, we no
longer focus on the extracted rationale candidate Z as the subject of study, but rather on the remaining
part of X after Z has been removed, denoted as X−Z . Regardless of whether the extractor selects S
or N to be the rationale candidate Z, we have
P (Y ∣X−Z ) = P (Y ∣X), s.t., Z ∈ {N, S}, (6)
The high level intuition behind Equation 6 is that neither removing the plain noise nor the spurious
features will cause a change in the task label. So, we have that
0 = DKL (P (Y ∣X−N )∣∣P (Y ∣X)) = DKL (P (Y ∣X−S )∣∣P (Y ∣X)) < DKL (P (Y ∣X−C )∣∣P (Y ∣X))
(7)
If we define the loss function as
L(Z) = −DKL (P (Y ∣X−Z )∣∣P (Y ∣X)), (8)
we will have that L(C) < L(N ) = L(S), which means that
L(N ) − L(C) L(S) − L(C)
g(N Ð
→ C) = > 0, g(S Ð
→ C) = > 0,
d(N, C) d(S, C)
(9)
L(N ) − L(S) L(S) − L(N )
g(N Ð
→ S) = = 0, g(S Ð
→ N) = = 0,
d(N, S) d(S, N )
where g(N Ð → C) is mentioned in the above qualitative analysis following Equation 5, denoting the
approximate tendency of the extractor to move from N to C. We call this objective as maximum
remaining discrepancy (MRD) criterion. The unique advantage of MRD is that it can treat spurious
features as equivalent to plain noise. Thus, extracting rationales from datasets containing spurious

6
 update the predictor to approximate the distribution

Predictor 𝑃𝑌෠ 𝑋 ȁ𝑋 𝑚𝑖𝑛(𝐻𝑐 (𝑌, 𝑌෠𝑋 |𝑋)+𝐻𝑐 (𝑌, 𝑌෠−𝑍 |𝑋−𝑧 ))

𝜃𝑝
Shared
𝑿 Extractor 𝑋−𝑍 Predictor 𝑃𝑌෠ −𝑍 ห𝑋−𝑧 𝑚𝑖𝑛 −DKL(𝑃𝑌෠ 𝑋 ȁ𝑋 ||𝑃𝑌෠ −𝑍 ห𝑋−𝑍 )
𝜃𝑒
𝑍
update the extractor for better rationale

Figure 4: The architecture of our proposed MRD. The approximators for the two distributions are
shared to reduce the model complexity.

features becomes equivalent to extracting from clean datasets without such features. As a result,
the extractor only needs to distinguish between noise and causal features, significantly reducing the
difficulty of rationale extraction.

5 The practical method

If we use Equation 8 to replace MMI, as long as C is not selected as the rationale candidate Z, the
objective will not distinguish between N and S. That is to say, no matter the extractor currently
selects N or S as the rationale candidate Z, the gradient descent algorithm can only guide it to move
towards C. It should be noted that the compactness of Z is facilitated through the sparsity constraint
expressed in Equation 4.
The followed problem is how to apply MRD in practice. The real distributions of P (Y ∣X−Z ) and
P (Y ∣X) are not directly accessible. So we need further efforts to approximate them. Similar to
the vanilla RNP’s approximating entropy with cross-entropy and inspired by the MCD’s (Liu et al.,
2023b) success in approximating real distributions with a predictor’s output, we try to approximate
real distributions by making use of the predictor. We first approximate P (Y ∣X) with P (ŶX ∣X)
by minimizing Hc (Y, ŶX ∣X) (please refer to Appendix A.7 for detailed analysis on the feasibility
of this approximation), and we also approximate P (Y ∣X−Z ) with P (Ŷ−Z ∣X−Z ) by minimizing the
cross-entropy Hc (Y, Ŷ−Z ∣X−Z ), where Ŷ−Z , ŶX are the predictor’s outputs with the inputs being
X−Z and X, respectively.
Finally, the training process for our MRD is depicted in Figure 4: the extractor first selects a rationale
candidate Z from the input X. Subsequently, X−Z and X are fed into the predictor to obtain two
distributions, P (Ŷ−Z ∣X−Z ) and P (ŶX ∣X). The overall objective of our model becomes (The pytorch
implementation is in Appendix A.8):
min[Hc (Y, ŶX ∣X) + Hc (Y, Ŷ−Z ∣X−Z )]
θp

+ min[−DKL (P (ŶX ∣X)∣∣P (Ŷ−Z ∣X−Z )) + Ω(M )], (10)

θe

s.t., (X, Y ) ∼ D, P (ŶX ∣X) = fP (X), X−Z = X − fE (X), P (Ŷ−Z ∣X−Z ) = fP (X−Z ),
where Ω(M ) is mentioned in Equation 4. The first term is used to help the predictor approximate the
distributions, and the second term helps the extractor find a good rationale.

6 Experiments
6.1 Datasets and metrics

Datasets. To validate the method’s ability to extract causal rationales in the input, there are certain
requirements for the datasets. First, the datasets should contain spurious correlations, making causality
a primary challenge within these datasets. Second, the test set should contain manually annotated
causal rationales to facilitate quantitative comparisons between different methods.
We employ six datasets collected from two widely used benchmarks. BeerAdvocate (McAuley et al.,
2012) is a benchmark that contains three widely used text classification datasets: Beer-Appearance,

7
Table 1: Results on Beer-Appearance and Beer-Aroma. Values in “()" are the standard deviations.
Datasets Beer-Appearance Beer-Aroma
Methods S P R F1 S P R F1
RNP 11.3 (0.7) 79.2 (4.4) 48.3 (0.9) 60.0 (1.2) 8.6 (0.6) 61.0 (29.4) 33.9 (17.2) 43.5 (21.7)
INVRAT 10.0 (n/a) 42.6 (0.7) 31.5 (0.6) 36.2 (0.6) 10.0 (n/a) 41.2 (0.3) 39.1 (2.8) 40.1 (1.6)
Inter_RAT 11.7 (0.6) 66.0 (0.4) 46.5 (0.8) 54.6 (0.7) 11.7 (0.6) 55.4 (0.9) 47.5 (0.6) 51.1 (0.8)
S ≈ 10%
NIR 11.0 (0.8) 79.8 (6.5) 47.1 (0.5) 59.2 (1.8) 10.3 (1.2) 72.1 (2.3) 47.6 (5.7) 57.2 (4.4)
MCD 9.5 (0.4) 94.2 (1.6) 48.4 (1.6) 63.9 (1.2) 9.9 (0.2) 84.6 (1.3) 53.9 (0.8) 65.8 (0.8)
MRD (ours) 10.0 (0.3) 93.6 (1.3) 50.7 (1.3) 65.7 (1.1) 10.1 (0.4) 86.6 (4.2) 56.2 (1.2) 68.1 (1.9)

RNP 20.5 (0.2) 70.0 (1.5) 77.4 (2.2) 73.5 (1.8) 19.4 (0.3) 61.0 (2.3) 76.0 (2.6) 67.7 (2.4)
INVRAT 20.0 (n/a) 58.9 (0.4) 67.2 (2.3) 62.8 (1.1) 20.0 (n/a) 29.3 (1.0) 52.1 (0.6) 37.5 (0.6)
Inter_RAT 21.7 (0.3) 62.0 (0.5) 76.7 (1.7) 68.6 (0.4) 20.4 (0.6) 44.2 (0.1) 65.4 (0.2) 52.8 (0.1)
S ≈ 20%
NIR 20.2 (0.7) 74.6 (4.4) 81.0 (2.0) 77.6 (3.2) 19.0 (0.2) 64.1 (1.6) 78.0 (1.2) 70.4 (1.4)
MCD 20.0 (0.3) 79.3 (0.6) 85.5 (1.1) 82.3 (0.5) 19.3 (0.2) 65.8 (0.7) 81.4 (1.3) 72.8 (0.9)
MRD (ours) 20.4 (0.5) 80.2 (2.3) 88.5 (1.0) 84.1 (1.5) 19.2 (0.4) 66.7 (1.3) 81.7 (1.8) 73.6 (1.3)

RNP 31.2 (1.0) 56.0 (2.0) 94.3 (1.5) 70.3 (1.9) 30.2 (0.8) 40.8 (4.1) 79.1 (8.0) 53.9 (5.4)
INVRAT 30.0 (n/a) 41.5 (0.4) 74.8 (0.3) 53.4 (0.3) 30.0 (n/a) 22.8 (1.6) 65.1 (1.7) 33.8 (1.8)
Inter_RAT 30.5 (1.0) 48.1 (0.7) 82.7 (0.4) 60.8 (0.4) 29.4 (0.6) 37.9 (0.7) 72.0 (0.1) 49.6 (0.7)
S ≈ 30%
NIR 29.6 (0.2) 59.6 (0.6) 95.3 (0.4) 73.3 (0.5) 29.6 (0.6) 43.3 (2.3) 82.4 (4.3) 56.8 (3.0)
MCD 29.7 (0.4) 59.6 (0.5) 95.6 (0.8) 73.4 (0.4) 29.6 (0.4) 46.1 (0.2) 87.5 (1.3) 60.4 (0.4)
MRD (ours) 28.6 (0.3) 60.6 (0.7) 93.3 (0.4) 73.5 (0.5) 29.3 (0.2) 46.8 (0.6) 88.3 (1.4) 61.2 (0.8)

Beer-Aroma, Beer-Palate. In these datasets, each piece of text is a comment consisting of the beer’s
three aspects: appearance, aroma, palate. And the comments of different aspects are highly correlated.
For the Beer-Appearance dataset, the classification label is the quality (bad/good, [0,1]) of the beer’s
appearance. Other two datasets are similar. These three datasets are most important and used by nearly
all of previous research in the field of rationalization. HotelReview (Wang et al., 2010) is a benchmark
that contains three widely used datasets: Hotel-Location, Hotel-Service, Hotel-Cleanliness. In these
datasets, each piece of text is a review about a hotel. For the Hotel-Location dataset, the classification
label is the quality (bad/good, [0,1]) of the hotel’s location. For Hotel-Service and Hotel-Cleanliness,
the classification label is about the service and cleanliness, respectively.
Metrics. Considering that the annotators assign the label of the target aspect by observing the causal
features, the overlap between the tokens selected by the model and those annotated by humans
provides a robust metric for rationale causality. The terms P, R, F 1 denote precision, recall, and
F 1 score respectively. These metrics are the most frequently used in rationalization. The term S
represents the average sparsity of the selected rationales, that is, the average percentage of selected
tokens in relation to the full text.

6.2 Baselines and implementation details

We compare with various recent methods to show the competitiveness of our method. These methods
include INVRAT (Chang et al., 2020), Inter_RAT (Yue et al., 2023), CR (Zhang et al., 2023), MCD
(Liu et al., 2023b), NIR (Storek et al., 2023). Both the extractor and the predictor are composed of an
encoder (e.g., RNN/Transformer) and a linear layer. We use two types of encoders: GRUs (following
INVRAT, Inter_RAT, and MCD, Table 1, 2, and 3) and bert-base-uncased (following CR, Table 4).
We adopt three levels of rationale sparsity: 10%, 20%, 30% (achieved by adjusting s in Equation 4).
We report the results of five random seeds. More details are in Appendix A.9.

6.3 Results

The main results2 are shown in Table 1, 2, and 3. Across various datasets and levels of rationale
sparsity, our proposed MRD achieves considerable improvements compared to existing baseline
methods. Compared to the most competitive baseline MCD, our MRD improves the F1 score by up
to 10.4% (=63.5% − 53.1%, in Beer-Palate dataset with S ≈ 10%). In addition, compared to the latest
penalty-based method Inter_RAT, we improve the F1 score by more than 10% in 14 out of 18 settings,
2
For the three beer-related datasets, the results of INVRAT and Inter_RAT are obtained from Table 1 of the
paper Inter_RAT. Since INVRAT requires specific techniques to partition datasets into environments, and it no
longer represents the latest literature, we have not replicated it on hotel-related datasets.

8
 Table 2: Results on Beer-Palate and Hotel-Location datasets.
Datasets Beer-Palate Hotel-Location
Methods S P R F1 S P R F1
RNP 10.1 (0.9) 58.5 (2.3) 47.6 (2.8) 52.4 (1.2) 9.9 (0.2) 47.9 (1.2) 55.6 (1.2) 51.4 (1.0)
INVRAT 10.0 (n/a) 34.9 (1.5) 45.6 (0.2) 39.5 (1.0) - - - -
Inter_RAT 12.6 (0.8) 34.6 (0.8) 48.2 (0.4) 40.2 (0.5) 11.8 (1.5) 31.6 (2.4) 43.2 (3.5) 36.4 (1.4)
S ≈ 10%
NIR 8.3 (3.3) 29.6 (20.0) 19.8 (17.7) 23.1 (18.6) 9.8 (0.6) 47.4 (1.6) 54.9 (1.9) 50.8 (1.0)
MCD 9.4 (0.8) 60.9 (2.1) 47.1 (3.0) 53.1 (1.9) 9.8 (0.3) 49.3 (2.1) 57.0 (3.0) 52.7 (2.4)
MRD (ours) 10.1 (0.3) 70.7 (2.0) 57.6 (2.1) 63.5 (1.9) 9.7 (0.2) 51.0 (1.6) 58.2 (1.6) 54.4 (1.6)

RNP 19.7 (0.2) 38.3 (1.6) 60.8 (3.1) 47.0 (2.1) 20.3 (0.4) 33.3 (1.0) 79.7 (2.5) 47.0 (1.4)
INVRAT 20.0 (n/a) 24.0 (1.3) 55.2 (2.3) 33.5 (1.6) - - - -
Inter_RAT 20.8 (0.6) 26.3 (0.6) 59.1 (0.8) 36.4 (0.7) 19.6 (1.4) 23.6 (0.7) 54.1 (2.6) 32.9 (0.4)
S ≈ 20%
NIR 19.5 (1.0) 32.9 (9.0) 51.8 (14.8) 42.0 (11.1) 20.0 (0.3) 33.0 (0.9) 77.6 (1.7) 46.3 (1.2)
MCD 19.6 (0.5) 41.2 (1.4) 65.0 (2.8) 50.5 (1.8) 19.7 (0.4) 33.8 (1.3) 78.5 (2.1) 47.3 (1.6)
MRD (ours) 19.6 (0.7) 44.2 (1.9) 69.6 (1.0) 54.1 (1.7) 19.4 (0.1) 35.0 (0.4) 79.5 (1.0) 48.6 (0.6)

RNP 29.1 (0.9) 24.2 (5.2) 56.7 (10.9) 34.0 (7.0) 29.5 (1.7) 18.1 (8.7) 64.2 (31.7) 28.2 (13.7)
INVRAT 20.0 (n/a) 20.9 (1.1) 71.6 (0.4) 32.3 (1.3) - - - -
Inter_RAT 30.4 (0.4) 21.8 (0.1) 66.1 (0.8) 32.8 (0.1) 29.8 (1.2) 18.1 (0.5) 63.1 (1.6) 28.1 (0.7)
S ≈ 30%
NIR 30.0 (3.7) 17.2 (8.6) 42.6 (22.4) 24.5 (12.4) 29.4 (0.9) 12.3 (10.6) 43.6 (37.6) 19.2 (16.6)
MCD 29.4 (1.7) 30.5 (1.0) 72.4 (5.6) 42.9 (1.8) 30.2 (0.3) 22.3 (1.8) 79.4(7.1) 34.8 (2.9)
MRD (ours) 28.2 (0.9) 30.9 (2.7) 70.3 (6.3) 43.0 (3.7) 29.4 (1.1) 25.4 (0.7) 88.0 (1.6) 39.5 (0.8)

Table 3: Results on Hotel-Service and Hotel-Cleanliness datasets.

Datasets Hotel-Service Hotel-Cleanliness
Methods S P R F1 S P R F1
RNP 10.1 (0.4) 46.1 (1.6) 40.4 (0.5) 43.1 (0.5) 9.8 (0.2) 33.8 (0.5) 37.6 (0.7) 35.6 (0.4)
Inter_RAT 11.2 (0.6) 32.6 (0.9) 32.3 (1.4) 32.4 (0.8) 9.4 (0.6) 32.5 (1.4) 34.5 (1.1) 33.4 (0.7)
S ≈ 10% NIR 10.7 (0.3) 44.8 (1.4) 41.9 (1.7) 43.3 (1.4) 10.2 (0.3) 35.1 (0.7) 40.5 (0.9) 37.6 (0.6)
MCD 10.2 (0.4) 47.5 (1.2) 42.3 (1.8) 44.7 (1.3) 9.8 (0.3) 34.3 (0.4) 37.8 (0.6) 35.9 (0.4)
MRD (ours) 10.5 (0.3) 48.5 (1.9) 44.3 (1.3) 46.3 (1.5) 9.9 (0.4) 34.6 (0.5) 38.8 (1.3) 36.6 (0.5)

RNP 20.0 (0.3) 31.8 (1.3) 55.4 (2.2) 40.4 (1.6) 20.7 (0.5) 21.5 (0.9) 50.3 (2.5) 30.1 (1.3)
Inter_RAT 20.6 (0.3) 24.5 (0.4) 44.7 (1.1) 31.7 (0.5) 19.5 (1.1) 22.7 (0.7) 50.1 (1.7) 31.3 (0.5)
S ≈ 20%
NIR 20.0 (0.5) 33.4 (0.7) 58.3 (0.5) 42.5 (0.5) 20.6 (0.5) 21.7 (0.5) 50.5 (1.0) 30.3 (0.6)
MCD 20.2 (0.3) 32.5 (0.5) 57.2 (1.4) 41.4 (0.7) 20.1 (0.5) 22.2 (0.5) 50.5 (1.4) 30.8 (0.7)
MRD (ours) 20.0 (0.6) 34.6 (1.4) 60.3 (1.1) 44.0 (1.4) 20.2 (1.4) 22.8 (0.6) 52.0 (2.2) 31.7 (0.3)

RNP 30.6 (0.7) 14.6 (8.2) 38.4 (21.5) 21.1 (11.9) 30.1 (0.5) 15.0 (1.6) 51.0 (5.3) 23.2 (2.4)
Inter_RAT 30.8 (1.0) 19.6 (0.3) 53.5 (1.9) 28.7 (0.5) 29.6 (1.2) 17.1 (0.5) 57.5 (1.0) 26.4 (0.6)
S ≈ 30%
NIR 30.1 (0.6) 19.3 (10.8) 50.3 (28.1) 27.9 (15.6) 30.8 (0.8) 16.4 (0.4) 57.0 (2.8) 25.4 (0.8)
MCD 30.1 (0.5) 22.5 (1.6) 59.0 (4.6) 32.5 (2.4) 30.2 (0.4) 16.5 (0.3) 56.3 (1.7) 25.5 (0.6)
MRD (ours) 30.1 (0.3) 24.7 (0.7) 64.9 (2.0) 35.8 (1.0) 29.2 (0.6) 18.8 (0.1) 62.1 (1.6) 28.9 (0.3)

and by more than 20% in 2 out of 18 settings, verifying the limitation of penalty-based methods. We
provide a visualized example of the extracted rationales by different methods in Appendix A.10.
We also follow a recent method CR (Zhang et al., 2023) to conduct experiments with the BERT
encoder as a supplement, whose results are shown in Table 4. We follow CR to set the sparsity level
as 10%, and the datasets are the most widely used Beer-Appearance and Beer-Aroma. Since some
methods become highly sensitive to hyperparameters after switching to an over-parameterized BERT
model (also supported by Remark 6.1 in (Zhang et al., 2023)), and our computational resources
are insufficient for extensive hyperparameter tuning for these methods, we primarily compare our
approach with methods that have already been implemented using BERT. Our MRD still outperforms
all the baselines. Specifically, we improve the F1 score by 15.6% on the Beer-Appearance dataset,
and 6.0% on the Beer-Aroma dataset.

7 Conclusion, limitations, and future work

This paper investigates the susceptibility of the widely adopted MMI criterion in XAI to spurious
correlations. We design a new criterion that can treat spurious features as plain noise, making rationale
extraction from datasets rich in spurious features as straightforward as extracting from clean datasets,
thus simplifying rationale extraction. Given the versatility of the self-explaining rationalization

9
Table 4: Results with BERT. We follow CR to set S ≈ 10%. ∗: results obtained from Table 11 of CR.
Datasets Beer-Appearance Beer-Aroma
Methods S P R F1 S P R F1
RNP∗ 10.0 (n/a) 40.0 (1.4) 20.3 (1.9) 25.2 (1.7) 10.0 (n/a) 49.1 (3.2) 28.7 (2.2) 32.0 (2.5)
VIB∗ 10.0 (n/a) 52.6 (2.0) 26.0 (2.3) 32.9 (2.1) 10.0 (n/a) 54.2 (2.9) 31.6 (1.9) 37.7 (2.8)
A2R∗ 10.0 (n/a) 55.0 (0.8) 25.8 (1.6) 34.3 (1.4) 10.0 (n/a) 61.3 (2.8) 34.8 (3.1) 41.2 (3.3)
S ≈ 10% ∗
INVRAT 10.0 (n/a) 56.4 (2.5) 27.3 (1.2) 36.7 (2.1) 10.0 (n/a) 49.6 (3.1) 27.5 (1.9) 33.2 (2.6)
∗
CR 10.0 (n/a) 59.7 (1.9) 31.6 (1.6) 39.0 (1.5) 10.0 (n/a) 68.0 (2.9) 42.0 (3.0) 49.1 (2.8)
MRD (ours) 10.6 (1.1) 75.0 (15.2) 43.0 (6.3) 54.6 (8.8) 9.9 (0.5) 71.7 (4.6) 44.8 (4.3) 55.1 (4.5)

framework, exploring how our method can be applied to broader fields such as computer vision and
graph learning is a worthwhile future direction.
One limitation is that, although some researchers have found that rationalization can benefit large
language models (LLMs) by providing high quality data (please refer to Appendix A.11), this paper
does not involve LLMs. Given the recent remarkable success of LLMs, exploring how our MRD can
aid in training trustworthy LLMs is another avenue worth pursuing.

8 Acknowledgements
This work is supported by the National Natural Science Foundation of China under grants 62376103,
62302184, 62436003 and 62206102; the Science and Technology Support Program of Hubei Province
under grant 2022BAA046; Hubei Science and Technology Talent Service Project under grant
2024DJC078; and Ant Group through CCF-Ant Research Fund.
We thank Lang Gao and Yang Qiu for their help during the rebuttal period. They helped a lot with the
experiments on LLMs and graph data.
We are also grateful for the valuable suggestions provided by the anonymous reviewers, which greatly
helped to improve the quality of this paper.

10
References
Julius Adebayo, Justin Gilmer, Michael Muelly, Ian J. Goodfellow, Moritz Hardt, and Been Kim.
2018. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems 31:
Annual Conference on Neural Information Processing Systems (NeurIPS 2018), December 3-8,
2018, Montréal, Canada, pages 9525–9536.

Martín Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. 2019. Invariant risk mini-
mization. CoRR, abs/1907.02893.

Christopher M. Bishop. 2006. Pattern Recognition and Machine Learning (Information Science and
Statistics). Springer-Verlag, Berlin, Heidelberg.

Aaron Chan, Maziar Sanjabi, Lambert Mathias, Liang Tan, Shaoliang Nie, Xiaochang Peng, Xiang
Ren, and Hamed Firooz. 2022. UNIREX: A unified learning framework for language model
rationale extraction. In International Conference on Machine Learning, ICML 2022, 17-23 July
2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research,
pages 2867–2889. PMLR.

Shiyu Chang, Yang Zhang, Mo Yu, and Tommi S. Jaakkola. 2020. Invariant rationalization. In
Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July
2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1448–1458.
PMLR.

Howard Chen, Jacqueline He, Karthik Narasimhan, and Danqi Chen. 2022. Can rationalization
improve robustness? In Proceedings of the 2022 Conference of the North American Chapter of the
Association for Computational Linguistics: Human Language Technologies, NAACL 2022, Seattle,
WA, United States, July 10-15, 2022, pages 3792–3805. Association for Computational Linguistics.

Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher,
and Byron C. Wallace. 2020. ERASER: A benchmark to evaluate rationalized NLP models. In
Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages
4443–4458, Online. Association for Computational Linguistics.

Marzyeh Ghassemi, Luke Oakden-Rayner, and Andrew L Beam. 2021. The false hope of cur-
rent approaches to explainable artificial intelligence in health care. The Lancet Digital Health,
3(11):e745–e750.

Shurui Gui, Meng Liu, Xiner Li, Youzhi Luo, and Shuiwang Ji. 2023. Joint learning of label and
environment causal independence for graph out-of-distribution generalization. arXiv preprint
arXiv:2306.01103.

Zishan Guo, Renren Jin, Chuang Liu, Yufei Huang, Dan Shi, Supryadi, Linhao Yu, Yan Liu, Jiaxuan
Li, Bojian Xiong, and Deyi Xiong. 2023. Evaluating large language models: A comprehensive
survey. CoRR, abs/2310.19736.

Shuaibo Hu and Kui Yu. 2024. Learning robust rationales for model explainability: A guidance-based
approach. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages
18243–18251.

Alon Jacovi and Yoav Goldberg. 2020. Towards faithfully interpretable NLP systems: How should
we define and evaluate faithfulness? In Proceedings of the 58th Annual Meeting of the Association
for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 4198–4205. Association
for Computational Linguistics.

Han Jiang, Junwen Duan, Zhe Qu, and Jianxin Wang. 2024. Mare: Multi-aspect rationale extractor
on unsupervised rationale extraction. arXiv preprint arXiv:2410.03531.

Tao Lei, Regina Barzilay, and Tommi S. Jaakkola. 2016. Rationalizing neural predictions. In
Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing,
EMNLP 2016, Austin, Texas, USA, November 1-4, 2016, pages 107–117. The Association for
Computational Linguistics.

11
Yong Lin, Hanze Dong, Hao Wang, and Tong Zhang. 2022a. Bayesian invariant risk minimization.
In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans,
LA, USA, June 18-24, 2022, pages 16000–16009. IEEE.

Yong Lin, Shengyu Zhu, Lu Tan, and Peng Cui. 2022b. Zin: When and how to learn invariance without
environment partition? Advances in Neural Information Processing Systems, 35:24529–24542.

Zachary C Lipton. 2018. The mythos of model interpretability: In machine learning, the concept of
interpretability is both important and slippery. Queue, 16(3):31–57.

Wei Liu, Haozhao Wang, Jun Wang, Zhiying Deng, Yuankai Zhang, Cheng Wang, and Ruixuan Li.
2024. Enhancing the rationale-input alignment for self-explaining rationalization. In 40th IEEE
International Conference on Data Engineering, ICDE 2024, Utrecht, The Netherlands, May 13-16,
2024, pages 2218–2230. IEEE.

Wei Liu, Haozhao Wang, Jun Wang, Ruixuan Li, Xinyang Li, Yuankai Zhang, and Yang Qiu. 2023a.
MGR: multi-generator based rationalization. In Proceedings of the 61st Annual Meeting of the
Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada,
July 9-14, 2023, pages 12771–12787. Association for Computational Linguistics.

Wei Liu, Haozhao Wang, Jun Wang, Ruixuan Li, Chao Yue, and YuanKai Zhang. 2022. Fr: Folded
rationalization with a unified encoder. In Advances in Neural Information Processing Systems,
volume 35. Curran Associates, Inc.

Wei Liu, Jun Wang, Haozhao Wang, Ruixuan Li, Zhiying Deng, Yuankai Zhang, and Yang Qiu.
2023b. D-separation for causal self-explanation. In Advances in Neural Information Processing
Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023,
New Orleans, LA, USA, December 10 - 16, 2023.

Wei Liu, Jun Wang, Haozhao Wang, Ruixuan Li, Yang Qiu, Yuankai Zhang, Jie Han, and Yixiong
Zou. 2023c. Decoupled rationalization with asymmetric learning rates: A flexible lipschitz restraint.
In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,
KDD 2023, Long Beach, CA, USA, August 6-10, 2023, pages 1535–1547. ACM.

Dongsheng Luo, Wei Cheng, Dongkuan Xu, Wenchao Yu, Bo Zong, Haifeng Chen, and Xiang Zhang.
2020. Parameterized explainer for graph neural network. In Advances in Neural Information
Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020,
NeurIPS 2020, December 6-12, 2020, virtual.

Julian J. McAuley, Jure Leskovec, and Dan Jurafsky. 2012. Learning attitudes and attributes from
multi-aspect reviews. In 12th IEEE International Conference on Data Mining, ICDM 2012,
Brussels, Belgium, December 10-13, 2012, pages 1020–1025. IEEE Computer Society.

Qihan Ren, Jiayang Gao, Wen Shen, and Quanshi Zhang. 2024. Where we have arrived in proving
the emergence of sparse interaction primitives in DNNs. In The Twelfth International Conference
on Learning Representations.

Marco Túlio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "why should I trust you?": Ex-
plaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17,
2016, pages 1135–1144. ACM.

Elan Rosenfeld, Pradeep Kumar Ravikumar, and Andrej Risteski. 2021. The risks of invariant risk
minimization. In 9th International Conference on Learning Representations, ICLR 2021, Virtual
Event, Austria, May 3-7, 2021. OpenReview.net.

Cynthia Rudin. 2019. Stop explaining black box machine learning models for high stakes decisions
and use interpretable models instead. Nature Machine Intelligence, 1(5):206–215.

Benjamin B Seiler. 2023. Applications of Cooperative Game Theory to Interpretable Machine

Learning. Ph.D. thesis, Stanford University.

12
Lei Sha, Oana-Maria Camburu, and Thomas Lukasiewicz. 2021. Learning from the best: Ratio-
nalizing predictions by adversarial information calibration. In Thirty-Fifth AAAI Conference on
Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Arti-
ficial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial
Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 13771–13779. AAAI Press.

Hua Shen, Tongshuang Wu, Wenbo Guo, and Ting-Hao Kenneth Huang. 2022. Are shortest rationales
the best explanations for human understanding? In Proceedings of the 60th Annual Meeting of the
Association for Computational Linguistics (Volume 2: Short Papers), ACL 2022, Dublin, Ireland,
May 22-27, 2022, pages 10–19. Association for Computational Linguistics.

Adam Storek, Melanie Subbiah, and Kathleen R. McKeown. 2023. Unsupervised selective ratio-
nalization with noise injection. In Proceedings of the 61st Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023,
pages 12647–12659. Association for Computational Linguistics.

Lichao Sun, Yue Huang, Haoran Wang, Siyuan Wu, Qihui Zhang, Chujie Gao, Yixin Huang, Wenhan
Lyu, Yixuan Zhang, Xiner Li, Zhengliang Liu, Yixin Liu, Yijue Wang, Zhikun Zhang, Bhavya
Kailkhura, Caiming Xiong, Chao Zhang, Chaowei Xiao, Chunyuan Li, Eric P. Xing, Furong
Huang, Hao Liu, Heng Ji, Hongyi Wang, Huan Zhang, Huaxiu Yao, Manolis Kellis, Marinka
Zitnik, Meng Jiang, Mohit Bansal, James Zou, Jian Pei, Jian Liu, Jianfeng Gao, Jiawei Han, Jieyu
Zhao, Jiliang Tang, Jindong Wang, John Mitchell, Kai Shu, Kaidi Xu, Kai-Wei Chang, Lifang
He, Lifu Huang, Michael Backes, Neil Zhenqiang Gong, Philip S. Yu, Pin-Yu Chen, Quanquan
Gu, Ran Xu, Rex Ying, Shuiwang Ji, Suman Jana, Tianlong Chen, Tianming Liu, Tianyi Zhou,
William Wang, Xiang Li, Xiangliang Zhang, Xiao Wang, Xing Xie, Xun Chen, Xuyu Wang, Yan
Liu, Yanfang Ye, Yinzhi Cao, and Yue Zhao. 2024. Trustllm: Trustworthiness in large language
models. CoRR, abs/2401.05561.

Hongning Wang, Yue Lu, and Chengxiang Zhai. 2010. Latent aspect rating analysis on review
text data: a rating regression approach. In Proceedings of the 16th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, July 25-28, 2010,
pages 783–792. ACM.

Yingxin Wu, Xiang Wang, An Zhang, Xiangnan He, and Tat-Seng Chua. 2022. Discovering
invariant rationales for graph neural networks. In The Tenth International Conference on Learning
Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.

Mengzhou Xia, Sadhika Malladi, Suchin Gururangan, Sanjeev Arora, and Danqi Chen. 2024. LESS:
selecting influential data for targeted instruction tuning. CoRR, abs/2402.04333.

Mo Yu, Shiyu Chang, Yang Zhang, and Tommi S. Jaakkola. 2019. Rethinking cooperative rationaliza-
tion: Introspective extraction and complement control. In Proceedings of the 2019 Conference on
Empirical Methods in Natural Language Processing and the 9th International Joint Conference on
Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019,
pages 4092–4101. Association for Computational Linguistics.

Mo Yu, Yang Zhang, Shiyu Chang, and Tommi S. Jaakkola. 2021. Understanding interlocking
dynamics of cooperative rationalization. In Advances in Neural Information Processing Systems
34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December
6-14, 2021, virtual, pages 12822–12835.

Hao Yuan, Lei Cai, Xia Hu, Jie Wang, and Shuiwang Ji. 2022. Interpreting image classifiers by
generating discrete masks. IEEE Transactions on Pattern Analysis and Machine Intelligence,
44(4):2019–2030.

Linan Yue, Qi Liu, Yichao Du, Li Wang, Weibo Gao, and Yanqing An. 2024. Towards faithful expla-
nations: Boosting rationalization with shortcuts discovery. In The Twelfth International Conference
on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net.

Linan Yue, Qi Liu, Li Wang, Yanqing An, Yichao Du, and Zhenya Huang. 2023. Interventional
rationalization.

13
Wenbo Zhang, Tong Wu, Yunlong Wang, Yong Cai, and Hengrui Cai. 2023. Towards trustworthy
explanation: On causal rationalization. In International Conference on Machine Learning, ICML
2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning
Research, pages 41715–41736. PMLR.
Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min,
Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen,
Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and
Ji-Rong Wen. 2023. A survey of large language models. CoRR, abs/2303.18223.
Yunxiao Zhao, Zhiqiang Wang, Xiaoli Li, Jiye Liang, and Ru Li. 2024. AGR: Reinforced causal
agent-guided self-explaining rationalization. In Proceedings of the 62nd Annual Meeting of the
Association for Computational Linguistics (Volume 2: Short Papers), pages 510–518, Bangkok,
Thailand. Association for Computational Linguistics.
Xiao Zhou, Yong Lin, Weizhong Zhang, and Tong Zhang. 2022. Sparse invariant risk minimization.
In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore,
Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 27222–27244.
PMLR.

14
A Appendix
A.1 The comparison between MCD and MRD

This paper is inspired by a previous paper MCD (Liu et al., 2023b). This part aims to clarify the
distinct contributions of our MRD.
We first note that the preliminaries of causality (§3.2) is provided by the paper of MCD. And the
contribution of the causality analysis does not belong to us.
Apart from this, the practical network architecture (Figure 4) may also look like that of MCD.
However, the core of this paper focuses on studying different optimization objectives. In fact, in
the field of rationalization, the network structures of many different methods are similar; the key
difference lies in the optimization objectives. This phenomenon is akin to research in the GAN
(Generative Adversarial Nets) field, where diverse approaches often share similar architectures but
differ primarily in their optimization strategies.
Our primary contribution is that the proposed MRD criterion allows the objective to treat the spurious
features equally as noise. To the best of our knowledge, this is the first research that can treat spurious
features as noise. And we do not need to coordinate the penalty term.
In MCD, the objective for rationale selection is
min DKL (P (Y ∣X)∣∣P (Y ∣Z)). (11)
θe

While in our MRD, it is

min −DKL (P (Y ∣X)∣∣P (Y ∣X−Z )), (12)
θe
where X−Z is the remaining part after removing the selected rationale candidate Z from the full input
X.
The research motivations behind MCD and MRD are quite distinct, approaching the problem from
opposite perspectives. MCD focuses on the properties that the selected rationale candidate Z should
satisfy. On the other hand, MRD examines the properties that X should exhibit after discarding Z,
emphasizing what remains in the input after the rationale is removed. This contrast highlights a
fundamental shift in how the problem of extracting meaningful information is addressed.
Aside from the motivations, the novelty of the practical method in this paper is also considerable.
Most existing research primarily focuses on the selected rationale as the main subject of study,
whereas this paper shifts attention to the unselected remaining parts. While some methods in the field
of explainable AI have also considered the unselected portions, their primary purpose has been to
achieve comprehensiveness, treating the unselected parts as supplements to the main content and still
requiring the balancing of multiple objectives (Yu et al., 2019). Moreover, these methods consider
the unselected parts not for achieving causality but for other aspects of interpretability. This paper is
novel in suggesting that focusing solely (i.e., completely through out the selected rationale candidate)
on the unselected remaining parts can effectively achieve causality, marking a distinctive approach in
the study of explainable AI.

A.2 A toy example of the backdoor path

This example is provided by (Liu et al., 2023b). To make the readers that are not familiar with causality
better understand the spurious correlations, we borrow it to provide a more intuitive understanding of
the correlation in Figure 2(b). We assume U , XA , XT , and Y are all Bernoulli variables, with their
respective probability distributions as:
p(U = 1) = p(U = 0) = 0.5,
p(XT = 1∣U = 1) = p(XT = 0∣U = 0) = 0.9,
(13)
p(XA = 1∣U = 1) = p(XA = 0∣U = 0) = 0.9,
p(Y = 1∣XA = 1) = p(Y = 0∣XA = 0) = 0.9.

With some simple derivations, we can easily obtain (detailed derivation is in Appendix A.4):
p(XA = 1) = p(XT = 1) = p(Y = 1) = 0.5. (14)

15
Then, we can further get (see Appendix A.5 for the detailed derivation of Equation 16 and 17):
p(U = 1, XT = 1) p(XT = 1∣U = 1)p(U = 1)
p(U = 1∣XT = 1) = = = 0.9. (15)
p(XT = 1) p(XT = 1)
p(XA = 1∣XT = 1) = ∑ p(XA = 1∣U )p(U ∣XT = 1) = 0.9 ∗ 0.9 + 0.1 ∗ 0.1 = 0.82. (16)
U ∈{0,1}

p(Y = 1∣XT = 1) = ∑ p(Y = 1∣XA )p(XA ∣XT = 1) = 0.82 ∗ 0.9 + 0.18 ∗ 0.1 = 0.756. (17)
XA ∈{0,1}

A.3 The association between different variables and Y

Though it is not the core claim of this paper, we will have a brief discussion about why LM M I (C) ≤
LM M I (S) < N .
The MMI loss is used to measure the indicative degree of Z towards the task label Y. First, we think
the noise N is independent of Y , thus it has the lowest mutual information with Y and the highest
MMI loss.
And for LM M I (C) ≤ LM M I (S), the reason is that C always co-occur with the target label in all
data samples. While in some data samples, there is not S but only C. So, C usually has higher
correlation with Y . This can also be understood from the probabilistic graphical model in Figure 2(a).
C is the direct cause of Y . The association between S and Y needs to flow through a path that passes
through C.

A.4 Derivation of Equation 14

We use XA as an example, and the others are nothing different.

p(XA = 1) = ∑ p(XA = 1, U ) = ∑ p(XA = 1∣U )p(U ) = 0.9 ∗ 0.5 + 0.1 ∗ 0.5 = 0.5. (18)
U ∈{0,1} U ∈{0,1}

A.5 Derivation of Equation 16 and 17

In Figure 2(b), we have XT á XA ∣U and XT á Y ∣XA . That is to say,

P (XA ∣U, XT ) = P (XA ∣U ), P (Y ∣XA , XT ) = P (Y ∣XA ). (19)

Then we can easily get Equation 16:

p(XA = 1∣XT = 1) = ∑ p(XA = 1, U ∣XT = 1)
U ∈{0,1}

= ∑ p(XA = 1∣U, XT = 1)p(U ∣XT = 1) (20)

U ∈{0,1}

= ∑ p(XA = 1∣U )p(U ∣XT = 1).

U ∈{0,1}

And Equation 17 is similar.

A.6 D-separation

D-separation is an important concept in probabilistic graphical models.

D-Separation (Bishop, 2006): A, B, and C denote arbitrary, non-intersecting sets of nodes (and
their union might not cover all nodes of the graph) in a given probabilistic graph. Our objective is to
determine whether a specific conditional independence statement A á B∣C is implied by this graph.
To do so, we examine all possible paths from any node in A to any node in B. A path is said to be
blocked if it includes a node o such that either

• (a) The arrows on the path meet at node o, forming either a chain (i.e., Ð
→oÐ
→) or a fork
ÐoÐ
(i.e., ← →), with the node o being part of set C, or

16
 • (b) The arrows on the path meet at node o to form a collider (i.e., Ð
→o←
Ð), and neither the
node o itself nor any of its descendants are included in set C.

If all paths are blocked, then A is considered to be d-separated from B by C, meaning that A á B∣C.
Liu et al. (2023b) have theoretically shown that Y and the non-causal features are d-separated by the
causal features.

A.7 Minimizing the cross-entropy is equal to minimizing the KL-divergence

The cross-entropy consists of two parts:

Hc (Y, ŶX ∣X) = H(Y ∣X) + DKL (P (Y ∣X)∣∣P (ŶX ∣X)). (21)

H(Y ∣X) is determined by the dataset itself and is irrelevant to the predictor. So, when we train a
predictor to minimize Hc (Y, ŶX ∣X), we are in fact minimizing DKL (P (Y ∣X)∣∣P (ŶX ∣X)). We
know that if and only if P (Y ∣X) = P (ŶX ∣X), we get the lowest KL-divergence (equal to 0).
So, we can finally use P (ŶX ∣X) to approximate P (Y ∣X) by training a predictor and minimizing
Hc (Y, ŶX ∣X).

A.8 The implementation with Pytorch

For a batch of (X, Y ), we first send X to the extractor to get Z and X−Z :

Z = fe (X), X−Z = X − Z. (22)

Then we get a copy of X−Z with the pytorch function “torch.detach()”:

′
X−Z = torch.detach(X−Z ). (23)

′
Then, we get ŶX and Ŷ−Z :
ŶX = fp (X),
′ ′
(24)
Ŷ−Z = fp (X−Z ).
Then we update the predictor with
′
min[torch.nn.functional.cross_entropy(Ŷ−Z , Y ) + torch.nn.functional.cross_entropy(ŶX , Y )],
θp
(25)
which is the first part of Equation 10. At the same time, we update the extractor with Equation 4.
Now, we deal with the second part of Equation 10. We first freeze the predictor’s parameters and get
X−Z again:
Z = fe (X), X−Z = X − Z. (26)

We now do not copy X−Z . Instead, we directly get ŶX and Ŷ−Z :

ŶX = fp (X),
(27)
Ŷ−Z = fp (X−Z ).

Then we update the extractor with

min −F.kl_div(F.softmax(Ŷ−Z ).log(), F.softmax(ŶX )), (28)

θe

where “F” denotes “nn.functional”. In practice, we have added Equation 4 to 28.

Now, an update round for Equation 10 is completed, and we repeat the above steps again.

17
 Table 5: Statistics of datasets used in this paper.
Train Dev Annotation
Datasets
Pos Neg Pos Neg Pos Neg Sparsity
Appearance 202385 12897 28488 1318 923 13 18.5
Beer Aroma 172299 30564 24494 3396 848 29 15.6
Palate 176038 27639 24837 3203 785 20 12.4
Location 7236 7236 906 906 104 96 8.5
Hotel Service 50742 50742 6344 6344 101 99 11.5
Cleanliness 75049 75049 9382 9382 99 101 8.9

A.9 More details

To the best of our knowledge, all datasets are sufficiently anonymized to make identification of
individuals impossible without significant effort. For beer-related datasets, users need to consult the
original authors (McAuley et al., 2012) for permission first.
There is another widely used version of BeerAdvocate where the data containing spurious correlations
has been manually removed by Lei et al. (2016). The cleaned version is used to study other problems
rather than causality. Since we are studying spurious correlations, we use the original version used by
Inter_RAT and MCD.
All datasets are in English. We process the datasets in the same way as MCD (Liu et al., 2023b).
The maximum text length is set to 256. More statistics of the datasets are in Table 5. The datasets
of BeerAdvocate is unbalanced. For the training data, we sample from the positive data to get same
number of positive and negative texts.
In practice, the approximators for the two distributions are shared to reduce model complexity. But
this trick is not necessary, if two separate nets are used to approximate the two distributions, the
performance can sometimes be even better.
Some previous methods needs very careful hyper-parameter tuning. To make fair comparisons, most
results of the baselines are copied from previous papers.
We follow MCD to use a learning rate of 0.0001 and a batchsize of 128 for the beer-related datasets.
For the hotel-related datasets, we also follow MCD to use a learning rate of 0.0001 and a batchsize of
256.
We report the average results of five different random seeds.
The experiments are run on a RTX4090 GPU, with 24GB memory.

A.10 Examples of the extracted rationales

We provide a visualized example of the rationales extracted by different methods in Figure 5. The
dataset is Beer-Appearance, and the rationale sparsity is set to about 10%. The causal rationale should
be the comments describing the beer’s appearance (the underlined texts). The vanilla RNP extracts
the taste as the rationale. Inter_RAT selects both aroma (“aroma is fruity”) and taste (“smooth and
very effervescent”). That is to say, both RNP and Inter_RAT select the spurious features as the
rationale. MCD selects both causal features (“yellow color ... notes”) and spurious features (“aroma
is fruity...”). While our MRD selects only the causal rationales.

A.11 The potential impact of rationalization in the era of LLMs

In comparison to traditional “model-centric" XAI methods which solely focus on the model’s learned
information, “data-centric" approaches primarily aim to extract model-agnostic patterns inherent in
the data. So, apart from improving interpretability, rationalization can serve as a method of data
cleaning (Seiler, 2023).
Domain-specific large models often require supervised fine-tuning using domain-specific data. Un-
cleaned data may contain harmful information such as biases and stereotypes (Sun et al., 2024).

18
 Label (Beer-Appearance): Positive. Label (Beer-Appearance): Positive.
Prediction: Positive. Prediction: Positive.
Input: this one was sent graciously to me by [unknown] , an Input: this one was sent graciously to me by [unknown] , an
awesome trader . cheers buddy yellow color beer with some awesome trader . cheers buddy yellow color beer with some
orange notes . head is white and disappear quickly . hazed and orange notes . head is white and disappear quickly . hazed and
very opaque . aroma is fruity ( cloves and bananas ) . typical of a very opaque . aroma is fruity ( cloves and bananas ) . typical of a
hefeweizen taste shows notes of orange and the typical hefeweizen taste shows notes of orange and the typical
hefeweizen taste ( cloves and bananas ) . smooth and very hefeweizen taste ( cloves and bananas ) . smooth and very
effervescent . almost no bitterness too . very drinkable and effervescent . almost no bitterness too . very drinkable and
refreshing . a nice hefeweizen refreshing . a nice hefeweizen

(a) RNP (b) Inter_RAT

Label (Beer-Appearance): Positive. Label (Beer-Appearance): Positive.

Prediction: Positive. Prediction: Positive.
Input: this one was sent graciously to me by [unknown] , an Input: this one was sent graciously to me by [unknown] , an
awesome trader . cheers buddy yellow color beer with some awesome trader . cheers buddy yellow color beer with some
orange notes . head is white and disappear quickly . hazed and orange notes . head is white and disappear quickly . hazed and
very opaque . aroma is fruity ( cloves and bananas ) . typical of a very opaque . aroma is fruity ( cloves and bananas ) . typical of a
hefeweizen taste shows notes of orange and the typical hefeweizen taste shows notes of orange and the typical
hefeweizen taste ( cloves and bananas ) . smooth and very hefeweizen taste ( cloves and bananas ) . smooth and very
effervescent . almost no bitterness too . very drinkable and effervescent . almost no bitterness too . very drinkable and
refreshing . a nice hefeweizen refreshing . a nice hefeweizen

(c) MCD (d) MRD (ours)

Figure 5: A visualized example of the rationales extracted by different methods.

Recent research suggests that training predictors with extracted rationales can remove irrelevant
harmful information, enhancing robustness (Chen et al., 2022) and generalization (Wu et al., 2022;
Gui et al., 2023).
Since LLMs are usually pretrained on various datasets, they tend to be less controllable than small
models (Zhao et al., 2023). Considering that for simple tasks (such as text classification), small
models are also capable and can achieve satisfactory results, we can train a separate rationalization
model for a single domain-specific dataset. Small models trained on a single dataset are often more
controllable and save computational resources (such as searching for hyperparameters and adding
regularization terms) (Guo et al., 2023). Then using the extracted rationales for supervised fine-
tuning might prevent large models from learning harmful information from new data. Additionally,
shortening input texts can also reduce the memory required for fine-tuning.
A recent study has also found that training a small model for data selection (although not the same as
rationale selection) and producing a small subset is useful for fine-tuning LLMs (Xia et al., 2024).

19
NeurIPS Paper Checklist
1. Claims
Question: Do the main claims made in the abstract and introduction accurately reflect the
paper’s contributions and scope?
Answer: [Yes]
Justification: [NA]
2. Limitations
Question: Does the paper discuss the limitations of the work performed by the authors?
Answer: [Yes]
Justification: Section 7.
3. Theory Assumptions and Proofs
Question: For each theoretical result, does the paper provide the full set of assumptions and
a complete (and correct) proof?
Answer: [Yes]
Justification: They are included in the Appendix.
4. Experimental Result Reproducibility
Question: Does the paper fully disclose all the information needed to reproduce the main ex-
perimental results of the paper to the extent that it affects the main claims and/or conclusions
of the paper (regardless of whether the code and data are provided or not)?
Answer: [Yes]
Justification: We provide the experimental details in Appendix A.9. We provide the Pytorch
implementation in Appendix A.8. We provide the code at https://anonymous.4open.
science/r/MRD-0427.
5. Open access to data and code
Question: Does the paper provide open access to the data and code, with sufficient instruc-
tions to faithfully reproduce the main experimental results, as described in supplemental
material?
Answer: [Yes]
Justification: We provide the code and data at https://anonymous.4open.science/r/
MRD-0427. We provide a detailed README.md to help the users use our code.
6. Experimental Setting/Details
Question: Does the paper specify all the training and test details (e.g., data splits, hyper-
parameters, how they were chosen, type of optimizer, etc.) necessary to understand the
results?
Answer: [Yes]
Justification: We provide them at Appendix A.9. Most of the Experimental Setting/Details
are borrowed from our baseline MCD.
7. Experiment Statistical Significance
Question: Does the paper report error bars suitably and correctly defined or other appropriate
information about the statistical significance of the experiments?
Answer: [Yes]
Justification: We report the results of five random seeds.
8. Experiments Compute Resources
Question: For each experiment, does the paper provide sufficient information on the com-
puter resources (type of compute workers, memory, time of execution) needed to reproduce
the experiments?
Answer: [Yes]

20
 Justification: They are included in Appendix A.9.
9. Code Of Ethics
Question: Does the research conducted in the paper conform, in every respect, with the
NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
Answer: [Yes]
Justification: There are no issues affecting the code of ethics in the study of this paper.
10. Broader Impacts
Question: Does the paper discuss both potential positive societal impacts and negative
societal impacts of the work performed?
Answer: [Yes]
Justification: This paper may inspire research in areas such as fairness and security. This
paper is unlikely to have a negative social impact.
11. Safeguards
Question: Does the paper describe safeguards that have been put in place for responsible
release of data or models that have a high risk for misuse (e.g., pretrained language models,
image generators, or scraped datasets)?
Answer: [NA]
Justification: [NA]
12. Licenses for existing assets
Question: Are the creators or original owners of assets (e.g., code, data, models), used in
the paper, properly credited and are the license and terms of use explicitly mentioned and
properly respected?
Answer: [Yes]
Justification: Appendix A.9.
13. New Assets
Question: Are new assets introduced in the paper well documented and is the documentation
provided alongside the assets?
Answer: [NA]
Justification: [NA]
14. Crowdsourcing and Research with Human Subjects
Question: For crowdsourcing experiments and research with human subjects, does the paper
include the full text of instructions given to participants and screenshots, if applicable, as
well as details about compensation (if any)?
Answer: [NA] [NA]
Justification: [NA]
15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human
Subjects
Question: Does the paper describe potential risks incurred by study participants, whether
such risks were disclosed to the subjects, and whether Institutional Review Board (IRB)
approvals (or an equivalent approval/review based on the requirements of your country or
institution) were obtained?
Answer: [NA]
Justification: [NA]

21