Quantifying the Plausibility of Context Reliance in Neural Machine Translation

Dear Reviewer kdfb,

Thank you for taking the time to engage actively with our submission! We appreciate your comments stating that our paper is “well written and the presentation is quite clear” and “shows an understanding of the current problems in the interpretability literature.” We are especially happy that you find our proposed approach ingenious and our findings relevant to interpretability research.

Application of PECoRe to other tasks/modalities

As you correctly state, our framework can be applied to other tasks or modalities to study the impact of context dependence beyond context-aware MT. We agree that showcasing such applications would strengthen the presentation of our method, and for this reason, we include a new appendix section in the revised submission (Appendix J). This new section applies PECoRe to a state-of-the-art decoder-only large language model to quantify context reliance in question answering, story generation tasks and information extraction. Our preliminary evidence suggests that PECoRe can be useful to ensure a trustworthy context reliance in these domains.

While such applications are undoubtedly exciting, we opt to include those as appendix materials since a more thorough evaluation of PECoRe for these tasks would, in our opinion, be more fitting for future studies leveraging task and modality-specific expertise to extend our investigation.

Statement conflating attribution plausibility with faithfulness

We agree that the sentence you report assumes a relation between faithfulness and plausibility when these two evaluation dimensions were shown to be largely independent in previous interpretability work. We have adjusted our formulation at the end of Section 4.5 to clarify that we are specifically interested in attribution methods that better align with human rationales.

Question regarding $\tilde y^{*}$ and $\hat y$

Even for cases where $\tilde y^{\*}$ is identical to $\hat y$, the probabilities associated with the tokens of the two sequences are likely to be different. Indeed, while both are produced from target prefix $\hat y_{<t}$, $\tilde y^{*}$ use non-contextual source inputs, while $\hat y$ uses contextual source inputs. If the probabilities of the same next token $\hat y_t$ are different, the result of the contrastive target function $f_{tgt}$ will also be different, leading to non-zero attribution scores over the context.

The only case in which these scores would be equal to zero is when the probability of the next token (or the distribution over the vocabulary, for KL-Divergence) is exactly the same with and without input context, i.e. if context is not affecting the generation process. We have added a new remark (3.2) to clarify this aspect.

Dear Reviewer DAmX,

Thank you for taking the time to engage actively with our submission! We are glad you consider the topic of our research “important for trustworthy AI.”

Hard-to-read with unclear takeaways, plausibility-specific terminology and French

• We have revised remarks in Section 3.2 and results in Sections 4.3 and 4.5 to make the logic and conclusions clearer.
• We have added an appendix section (Appendix A) to introduce some terminology adopted in plausibility studies, provide a concrete example of current practices in plausibility evaluation, and clarify the functioning of our adopted plausibility metrics for people less familiar with this research domain.
• While it is inevitable to depend on non-English examples for a study focusing on machine translation, we added English glosses to Section 5 examples and appendix tables to simplify understanding of relevant elements in model translations.

Narrow scope, encoder-decoder specific method

The PECoRe framework is not architecture-specific and can be applied to modern decoder-only large language models. We apologize for not making this clearer in the initial version of the paper, and we have added a clarifying remark in the revised version of the introduction. Moreover, we have added some examples in Appendix J to further highlight the broad applicability of our approach to study the plausibility of a state-of-the-art decoder-only large language model to quantify context reliance in question answering, story generation tasks and information extraction. Our preliminary evidence suggests that PECoRe can be useful to ensure a trustworthy context reliance in these domains.

Lack of comparisons to established practices in plausibility evaluation

The established practice for plausibility evaluation is to verify the overlap between model rationales and human annotation for pre-defined target tokens of interest. Our framework, instead, proposes a more comprehensive evaluation framework that includes the identification of those targets. The comparison between the two is performed in Section 4.5. The CCI results from gold context-sensitive tokens in Figure 4 reflect the traditional hypothesis-driven procedure for plausibility evaluation. Additional background and a concrete example of plausibility evaluation are now provided in Appendix A to avoid any confusion.

Instead, end-to-end medians in the same Figure represent the outcome of our proposed framework, leading to more conservative plausibility estimates that reflect a broader notion of model contextual reliance, incorporating the detection of context-sensitive target tokens in the evaluation.

Importantly, the goal of our work is not to adopt the traditional “SOTA” procedure for plausibility evaluation to improve results on some metrics of interest (e.g. Macro F1 between human and model rationales in Figure 4), but to address the shortcomings of this procedure by proposing a shift from hypothesis-driven to data-driven plausibility evaluation. This shift ultimately enables its usage in real-world settings where references and annotations are unavailable. We added a sentence in the revised Section 4.5 to clarify this aspect.

In terms of our choice of methods, the P-CXMI method can be seen as the “previous SOTA” for CTI, as it was used by [1] for context sensitivity detection, while for CCI, the Attention Mean method by [2] was the best one that does not require pre-hoc knowledge (as in the case of Attention Best) before conducting the evaluation.

References

[1] When Does Translation Require Context? A Data-driven, Multilingual Exploration (Fernandes et al., ACL 2023)

[2] When and Why is Document-level Context Useful in Neural Machine Translation? (Kim et al., DiscoMT 2019)

Dear Reviewer GSke,

Thank you for taking the time to actively engage with our submission! We are glad you found our method “simple but sound, with easy-to-interpret outputs,” and appreciate your comment about our work addressing “an important problem in context-aware machine translation.”

Usage of mBART in Section 5 end-to-end experiments

Our initial decision to use mBART for our end-to-end analysis in Section 5 was aimed at showcasing the applicability of our approach to other models and target languages. However, we agree with your comment suggesting that missing OpusMT examples might undermine the systematicity and completeness of our analysis since the system was used in previous sections. For this reason, in the revised version of our paper, we extend our end-to-end analysis to the OpusMT Large model, adding two examples of OpusMT results in Section 5 and several more in Appendix I (Table 7). We observe that OpusMT results present several issues with lexical cohesion that closely match those previously shown for mBART, making the new result consistent with our original analysis.

Dear Reviewer sS3A,

Thank you for actively engaging with our submission! We appreciate your comments mentioning that “the topic is very interesting” and that “the evaluation encompasses a wide range of models.” We are especially glad you found our explanation of the PECoRe approach comprehensive.

Novelty and advantages of a unified CTI+CCI framework

While it is true that CTI and CCI can be applied independently, we want to emphasize that PECoRe is not a simple sequential application of existing approaches. On the one hand, current evaluation practices use non-contrastive attribution methods as the only step to quantify model plausibility. In those, the hypothesis-driven pre-selection of attribution targets in a hypothesis-driven way inevitably biases the evaluation to a subset of known and expected behaviors (e.g. we added a clarifying example in Appendix A). On the other hand, CTI metrics were used for other purposes, such as filtering datasets of examples with context-dependent phenomena [1], and were not applied in the context of plausibility evaluation in previous work.

Apart from selecting attribution targets in a data-driven way with CTI, PECoRe also uses CTI contextual and non-contextual outputs as targets for contrastive attribution, effectively harmonizing CTI inputs with CCI targets. To our knowledge, this is the first application of CTI methods and contrastive attribution for studying model plausibility, proposing a shift from hypothesis-driven to data-driven plausibility evaluation to enable its usage in real-world settings where references and annotations are unavailable (as we do in Section 5). Notably, [2] found the contrastive attribution approach we adopt for CCI to lead to improved simulatability for model decisions compared to regular attribution, providing an additional justification for our novel framework.

Limited performance and unclear generalizability to other contextual phenomena

Your statement, “the proposed methodology predominantly excels in handling anaphora cases,” implies that the weaker CTI scores on lexical choice in Figure 3 could be attributed to issues with our proposed PECoRe framework. However, this is not the case: Table 1 reports the performance of the MT model we use as a “test subject” and not the results of our proposed plausibility evaluation framework (Figures 3 and 4). These two aspects should not be conflated, as our method aims to detect cases of implausible model behavior and not to directly improve the performance of the underlying model.

Indeed, from Table 1, it is evident that the MT model can less proficiently use contextual information for lexical choice, likely due to the limited presence of such phenomena in the training data. If the model cannot use contextual information to disambiguate lexical choices in several cases (Table 1, column “OK”), it is expected that CTI would identify this setting as implausible from a context-sensitivity perspective (Figure 3, Left). This motivates us to include results for the OK-CS split (Figure 3, right), the data subset for which we know context availability leads to the correct disambiguation. On this split, we can see that the proposed KL-Divergence metric fares competitively, suggesting the robustness of PECoRe despite the domain shift.

As additional evidence of PECoRe’s generalizability to unseen phenomena, Appendix G (Table 5) in the revised draft presents some SCAT+ examples for whose a context sensitivity in the choice of grammatical formality (“you” → “toi/vous”) was correctly identified by the KL-Divergence CTI metric. However, those examples were considered wrong in the evaluation of Section 4.3 since it did not match the gender disambiguation phenomena (“it, they” → “il/elle, ils/elles”) used to annotate SCAT+.

Page 4 typos: We thank you for reporting these typos. Following other reviewers’ questions, Remark 3.1 is adjusted, and a new Remark 3.2 is added in the revised version to further clarify the case in which the presence of input context leads to the same generations for $\tilde y^{\*}$.

References

[1] When Does Translation Require Context? A Data-driven, Multilingual Exploration (Fernandes et al., ACL 2023)

[2] Interpreting Language Models with Contrastive Explanations. (Yin and Neubig, EMNLP 2022)