Towards Understanding Neural Machine Translation with Word Importance

Longyue Wang; Michael R. Lyu; Shilin He; Shuming Shi; Xing Wang; Zhaopeng Tu

REVIEW 4 major objections 5 minor 1 cited by

Towards Understanding Neural Machine Translation with Word Importance

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that integrated gradients, a gradient-based attribution method, consistently outperforms attention and black-box baselines at estimating which source words matter to a neural machine translation model, as shown by…

desk verdict A solid empirical comparison showing gradient-based attribution beats attention for NMT word importance; the BLEU-degradation evaluation is imperfect but the central ranking claim holds up. read the letter →

arxiv 1909.00326 v2 pith:CGSG5L6F submitted 2019-09-01 cs.CL cs.LG

Shilin He , Zhaopeng Tu , Xing Wang , Longyue Wang , Michael R. Lyu , Shuming Shi This is my paper · ORCID

classification cs.CLcs.LG

keywords neuralmachinetranslationintegratedgradientswordimportanceattentioninterpretabilityunder-translationlanguage-pairdifferences

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a gradient-based attribution method, integrated gradients, gives a more reliable estimate of which source words actually influence a neural machine translation (NMT) model's output than attention scores or frequency/content heuristics. The authors validate this by perturbing the top-ranked words and measuring BLEU degradation across three perturbation types, three language pairs and their reverse directions, and two model architectures. In every setting, the attribution method drops translation quality the most, and it also flags human-annotated under-translated words more accurately than attention or erasure. If correct, the method offers an unsupervised, architecture-agnostic tool for debugging NMT and for studying how different languages place weight on different word types, which could inform architecture design.

What carries the argument

The central object is integrated gradients (IG), a white-box attribution method defined as $IG_n^m(x) = (x_m - x'_m) \int_0^1 \frac{\partial F(x' + \alpha(x-x'))_n}{\partial x_m} d\alpha$, where $x'$ is a zero-embedding baseline and $F(x)_n$ is the model's output probability for the $n$-th target word. In practice the integral is approximated by summing gradient evaluations over $S=300$ evenly spaced steps along the path. For each input word, contributions are aggregated across all output words by summation and normalized with softmax to yield a single importance score. This mechanism supplies a per-word ranking that is architecture-invariant and requires no ground-truth reference, and it is the component that the paper compares against attention and black-box baselines.

What would settle it

A control experiment would train an NMT model with a randomly initialized and frozen encoder so the source words carry no learned meaning; if the integrated-gradients ranking on that model still produces large BLEU drops under grammatical-replacement perturbation, then the evaluation proxy is measuring the perturbation itself rather than the word's learned influence.

Watch

Extended reading notes

Core claim

The central claim is that word importance estimated by integrated gradients—the integral of output-probability gradients with respect to input word embeddings along a path from a zero baseline—consistently outperforms attention, content-word, frequency, and random baselines under deletion, masking, and grammatical-replacement perturbations. The evaluation measure is the drop in BLEU when the top-ranked words are perturbed; the larger the drop, the more important the words are. The paper reports that attribution achieves the largest drops on Chinese-English, English-French, and English-Japanese (and their reverse directions) using both Transformer and RNN-Search models, that it outperforms the supervised erasure method once two or more words are perturbed, and that its importance scores detect under-translated words better than attention or erasure. The paper also reports that the word categories receiving the highest importance shift with the language pair—nouns for Chinese-English, prepositions for English-French, and prepositions and punctuation for English-Japanese—while high-fertility (one-to-many) words are consistently important and null-aligned words consistently unimportant.

Load-bearing premise

The evaluation equates word importance with the drop in BLEU when the word is perturbed, so side effects of the perturbation itself (such as sentence shortening or grammatical nonsense) can count as importance even if the word's intrinsic contribution is small.

Editorial extensions

If this is right

Gradient-based attribution can replace attention as the default way to locate influential input words when debugging or explaining an NMT model.
Low-importance words can be surfaced automatically as candidates for under-translation errors, supporting post-editing or constrained decoding without human annotation.
The language-specific importance patterns imply that NMT architectures should embed inductive biases tuned to the target language pair rather than assuming all languages weight word categories identically.
Because the method needs only gradients from a trained model, it transfers across Transformer and RNN-Search architectures and across language directions without retraining.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The evaluation proxy (BLEU drop under perturbation) conflates a word's intrinsic contribution with side effects of the perturbation itself, such as sentence shortening under deletion or semantic nonsense under grammatical replacement; if a future study used meaning-preserving substitutions instead, the ranking advantage might shrink.
The same integrated-gradient machinery could be applied to other conditional text generation tasks, such as summarization or data-to-text, to identify input tokens that drive output choices, provided a comparable perturbation-based evaluation is constructed.
The cross-language shift in important categories (nouns vs. prepositions vs. punctuation) suggests that transfer of attribution models across languages is not straightforward and that a single global notion of 'important word' may not exist.
One testable extension: use the importance scores as salience weights in training loss or data augmentation; if the paper's ranking is faithful, up-weighting high-importance words (or masking low-importance ones) should improve translation quality or robustness.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this paper applies integrated gradients to NMT word importance and shows, across three language pairs, two architectures, and three perturbation types, that it identifies words whose removal or corruption hurts translation performance more than attention, frequency, content, or random selection. That is a solid, useful result. The BLEU-degradation proxy is the main thing to quibble with, but it is not the fatal flaw the stress-test makes it out to be. The paper explicitly frames word importance by influence on translation output, and BLEU degradation is a natural measure of that. The Random baseline controls for the average effect of perturbation, so the comparative ranking is meaningful.

The genuinely new parts: applying IG to NMT word importance with this breadth of comparison, and the finding that Attribution beats the supervised Erasure baseline when multiple words are perturbed. That is a real point about multi-word interactions, and it cuts against the worry that the method is merely optimizing the BLEU criterion. The under-translation detection experiment, with modest F1 (0.25–0.34) but consistent gains over attention and erasure, gives some external anchor beyond the perturbation metric.

Where I would push back: the main BLEU curves have no error bars or significance tests. The 1.0+ BLEU gap between Attribution and Attention is consistent across all settings, so I am fairly convinced it is real, but the paper should report variance and ideally run a paired test. The under-translation annotation has ten annotators but no inter-annotator agreement, which matters for a subjective label. And the POS/fertility analysis is correlational; the architecture-design recommendation is speculative, and the paper mostly acknowledges that.

The stress-test note worries that BLEU degradation is the sole operationalization. That is fair, but the paper never claims to measure intrinsic linguistic importance—it measures influence on translation performance. Deletion and mask perturbations trigger brevity penalties and distributional shifts, but those effects apply equally to all methods, and the Random baseline absorbs them. So the central comparison survives. The one caveat I would add is that BLEU-based evaluation can only rank methods relative to each other; it does not tell us whether any of these scores correspond to what a human would call important. The under-translation task is a step in that direction, but only a small one.

No code is released, which is a real downside for reproducibility. The IG step count S=300 is described as chosen from preliminary experiments and not tuned to the target metric, so it is not circular, but the choice should be documented more carefully.

Who is this for? NMT interpretability researchers, and anyone choosing an attribution method for sequence-to-sequence models. It deserves a serious referee—revision, not desk reject. I would want significance tests, code release, and a bit more care around the BLEU proxy before signing off, but the core result is believable and useful.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes using integrated gradients (IG) to assign importance scores to source words in NMT, and evaluates the scores by measuring BLEU degradation when the top-scored words are deleted, masked, or replaced with same-POS words. It reports that this Attribution method outperforms black-box baselines (Content, Frequency) and Attention across three language pairs, both translation directions, and two architectures, and that it also identifies under-translated words better than Erasure and Attention on a small human-annotated set. It further analyzes POS and fertility correlates of importance, concluding that important syntactic categories vary across language pairs.

Significance. The paper addresses a real interpretability need in NMT and has three notable strengths: it uses an axiomatic attribution method (IG) with no parameters fitted to the evaluation metric; it covers a broad experimental matrix (three perturbations, three language pairs plus reverse directions, Transformer and RNN-Search); and it provides a human-annotated under-translation detection task that grounds the usefulness of the scores outside the perturbation framework. If the superiority claim is established, the method would be a practical tool for debugging and for designing language-pair-specific architectures. However, the evaluation rests almost entirely on an unvalidated BLEU-degradation proxy, and the headline comparisons lack statistical support. The human-grounded validation is narrow and modest in performance, so the external-validity case is currently incomplete.

major comments (4)

[Section 4, 'Evaluation'] The paper defines word importance operationally as BLEU degradation under perturbation ('The more translation performance degrades, the more important the word is'), but this proxy is not validated against any independent notion of causal importance. Deletion and Mask change source length or embedding statistics, so the resulting BLEU drop can reflect brevity penalties and distributional shifts rather than the intrinsic role of the selected word; Grammatical Replacement produces semantically incoherent inputs that can degrade BLEU for reasons unrelated to the target word. Because every headline comparison in Figures 2-4 is scored with this metric, the authors should provide evidence that BLEU-damage rankings correlate with human judgments of word importance, for example by extending the Section 5.1 annotation to a ranking task or by reporting a perturbation metric that controls for length (e.g., sentence-level semantic similarity or reference-free quality measures).
[Section 4.1, Figures 2-4] The central claim that Attribution 'consistently outperforms' other methods is made without error bars, confidence intervals, or significance tests. Only Random and Content are averaged over 10 runs; Attention and Attribution are deterministic, but the test set is finite and the reported Attribution-Attention gap is sometimes around 1 BLEU, which may be within test-set noise. A paired bootstrap or per-sentence significance test (e.g., 95% confidence intervals on the BLEU difference) is needed to support the superiority claim across operation counts and language pairs.
[Section 4.3, Figure 4] The comparison with Erasure is informative but does not validate the BLEU-degradation criterion itself: Erasure defines importance directly as per-word BLEU degradation, so the fact that Attribution outperforms Erasure on multi-word perturbations indicates how multi-word attribution aggregates under this criterion, not that the criterion measures true word importance. The manuscript should state this limitation, and ideally add a criterion-independent evaluation (e.g., human-ranked importance or performance on a downstream task such as under-translation detection) for the multi-word setting.
[Section 5.1, Table 1] The only human-grounded validation, under-translation detection, reports modest F1 scores (0.248-0.342), uses a small dataset (500 sentences, 178 with errors), provides no inter-annotator agreement measure, and does not test significance despite claiming that Attribution 'consistently and significantly outperforms' Erasure and Attention. Because the paper uses this experiment as external support for the attribution scores, the analysis should include agreement statistics and confidence intervals for the F1 differences.

minor comments (5)

[Section 3.3, Eq. (2)] The Riemann-sum approximation sums k=0 to S, which uses S+1 points and includes the baseline and the input endpoints; standard IG uses k=1 to S (or the average of adjacent terms for a trapezoid rule). Please clarify the discretization and its effect on the 300-step approximation.
[Section 4, Figure 2] Figure 2 is difficult to read: the three rows are labeled with the same legend and the (a)-(c) subcaptions appear to be misaligned with the perturbation types. Reorganize the panels and make the perturbation type clear in each subplot.
[Section 5.2, Table 2] The 'decision tree based regression' used to compute correlations is not described (e.g., tree depth, number of features, whether importance is used as the target); without this detail, the reported correlations in Table 2 are hard to interpret.
[Throughout] There are several typos and grammatical issues ('embeded', 'consists with', 'contributon', 'translated sentence' vs 'translation'); a careful proofread is needed.
[Section 5.2] The claim that 'we multiply the word importance with the corresponding sentence length for fair comparison' is not justified; please explain why this normalization is appropriate and whether results are robust to it.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: integrated-gradient importance scores are generated independently and validated against external perturbation BLEU changes and human under-translation labels, with no fitted parameter tied to the target metric.

full rationale

The paper's central derivation is self-contained rather than circular. Word importance is computed by integrated gradients (Section 3.3) from the NMT model's output probabilities and input embeddings, with no parameter fitted to BLEU or to the human labels. The evaluation (Section 4) then asks whether perturbing the top-scored words degrades translation performance more than perturbing random or baseline-selected words; this is an empirical validation of the attribution scores, not a derivation of those scores from the evaluation metric. The S=300 integration-step choice is described as a preliminary experimental setting, not as tuning against the downstream BLEU-degradation criterion. The Erasure baseline in Section 4.3 is defined directly in terms of BLEU-score change, but it is used as a comparative baseline rather than as the ground truth that defines Attribution, so the comparison does not reduce Attribution's claim to its own evaluation. The under-translation analysis (Section 5.1) uses independent human annotations of under-translated words, providing external validation rather than circular support. Citations to prior work such as Sundararajan et al. (2017) for integrated gradients and Jain and Wallace (2019) for attention limitations are external and not self-referential; the one self-citation (Yang et al., 2019) appears only in related work and is not load-bearing for any claim. The BLEU-degradation proxy is a legitimate external-validity concern about what 'importance' means, but it is not a circularity: the paper does not define its attribution scores in terms of BLEU, nor does it fit any parameter to the metric it predicts. Therefore the central comparison is not forced by construction.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The work rests on standard attribution axioms from prior work and on domain assumptions about BLEU as a fidelity metric, zero embeddings as a neutral baseline, and the reliability of human annotation. No new entities are introduced.

free parameters (1)

IG step count S = 300
Number of integration steps in the Riemann approximation of the integrated gradients integral; authors say preliminary experiments found 300 steps to give fairly good performance, but no systematic tuning on the target metric is reported.

assumptions (4)

domain assumption BLEU degradation under perturbation is a valid measure of word importance.
Section 4, Evaluation: 'The more translation performance degrades, the more important the word is.' This is a modeling assumption because BLEU drops can reflect length changes, POS replacement artifacts, and semantic incoherence beyond the targeted word.
standard math Integrated gradients satisfies the attribution axioms of sensitivity and implementation invariance as proven by Sundararajan et al., 2017.
The method inherits its theoretical justification from the cited work, invoked in Section 3.3.
domain assumption The zero-embedding sequence is a valid baseline for text attribution.
Zero embeddings are out-of-distribution for a trained NMT model; whether they represent a neutral or meaningful baseline for text is assumed, not justified.
domain assumption Human annotations of under-translation are reliable without reported inter-annotator agreement.
Section 5.1 describes 10 annotators labeling 500 sentence pairs, each pair labeled by at least two annotators, but no agreement statistic (e.g., Cohen's kappa) is reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Understanding Neural Machine Translation with Word Importance." pith.science (2026). https://pith.science/paper/CGSG5L6F

@misc{pith2026190900326,
  author       = {Pith},
  title        = {Pith review of: Towards Understanding Neural Machine Translation with Word Importance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CGSG5L6F}},
  note         = {Machine review of arXiv:1909.00326}
}

read the original abstract

Although neural machine translation (NMT) has advanced the state-of-the-art on various language pairs, the interpretability of NMT remains unsatisfactory. In this work, we propose to address this gap by focusing on understanding the input-output behavior of NMT models. Specifically, we measure the word importance by attributing the NMT output to every input word through a gradient-based method. We validate the approach on a couple of perturbation operations, language pairs, and model architectures, demonstrating its superiority on identifying input words with higher influence on translation performance. Encouragingly, the calculated importance can serve as indicators of input words that are under-translated by NMT models. Furthermore, our analysis reveals that words of certain syntactic categories have higher importance while the categories vary across language pairs, which can inspire better design principles of NMT architectures for multi-lingual translation.

Figures

Figures reproduced from arXiv: 1909.00326 by the authors.

**Figure 1.** An example of (a) word importance and (b) contribution matrix calculated by Attribution (integrated gradients) on English⇒French translation task. Input in English: “It has always taken place .” Output in French: “Elle a toujours eu lieu .” where ∂F(x)n ∂xm is the gradient of F(x)n w.r.t. the embedding of the mth word. In this paper, as suggested, the baseline input x 0 is set as a sequence of zero embeddings that… view at source ↗

**Figure 2.** Effect of three types of synthetic perturbations on Chinese [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Effect of the Mask perturbation on (a) Chinese⇒English translation using the RNN-Search model, (b, c, d, e, f) other language pairs and directions using Transformer model. Different Language Pairs and Directions We further conduct experiments on another two language pairs (i.e., English⇒French, English⇒Japanese in Figures 3(b, c)) as well as the reverse directions (Figures 3(d, e, f)) using Transformer under the Mas… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Effect of Attribution and Erasure methods on Chinese⇒English translation with Mask perturbation. In [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

The Role of Handling Attributive Nouns in Improving Chinese-To-English Machine Translation
cs.CL 2024-12 reject novelty 4.0 of 10

Manually inserting the omitted Chinese particle 的 into news titles and fine-tuning NMT models yields small BLEU and CHRF improvements, but the experiment has serious baseline and reporting gaps.

Reference graph

Works this paper leans on

34 extracted references · 26 canonical work pages · cited by 1 Pith paper

[1]

URL: " 'urlintro :=

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

David Alvarez-Melis and Tommi Jaakkola. 2017. A causal framework for explaining the predictions of black-box sequence-to-sequence models. In EMNLP

work page 2017
[4]

Leila Arras, Franziska Horn, Gr \'e goire Montavon, Klaus-Robert M \"u ller, and Wojciech Samek. 2016. Explaining predictions of non-linear classifiers in nlp. In Proceedings of the 1st Workshop on Representation Learning for NLP

work page 2016
[5]

Alexei Baevski and Michael Auli. 2019. Adaptive input representations for neural language modeling. In ICLR

work page 2019
[6]

Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. In ICLR

work page 2014
[7]

Anthony Bau, Yonatan Belinkov, Hassan Sajjad, Nadir Durrani, Fahim Dalvi, and James Glass. 2019. Identifying and controlling important neurons in neural machine translation. In ICLR

work page 2019
[8]

Yonatan Belinkov, Nadir Durrani, Fahim Dalvi, Hassan Sajjad, and James Glass. 2017. What do neural machine translation models learn about morphology? In ACL

work page 2017

Show all 34 references

[9]

Noam Chomsky and David W Lightfoot. 2002. Syntactic structures. Walter de Gruyter

2002
[10]

Trevor Cohn and Mirella Lapata. 2008. Sentence compression beyond word deletion. In COLING

2008
[11]

Kedar Dhamdhere, Mukund Sundararajan, and Qiqi Yan. 2019. How important is a neuron? In ICLR

2019
[12]

Yanzhuo Ding, Yang Liu, Huanbo Luan, and Maosong Sun. 2017. Visualizing and understanding neural machine translation. In ACL

2017
[13]

George Foster, Pierre Isabelle, and Pierre Plamondon. 1997. Target-text mediated interactive machine translation. Machine Translation, 12(1/2):175--194

1997
[14]

Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. 2017. Convolutional sequence to sequence learning. In ICML

2017
[15]

\'E douard Grave, Armand Joulin, Moustapha Ciss \'e , David Grangier, and Herv \'e J \'e gou. 2017. Efficient softmax approximation for GPU s. In ICML

2017
[16]

Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In NAACL

2018
[17]

Hany Hassan, Anthony Aue, Chang Chen, Vishal Chowdhary, Jonathan Clark, Christian Federmann, Xuedong Huang, Marcin Junczys-Dowmunt, William Lewis, Mu Li, et al. 2018. Achieving human parity on automatic chinese to english news translation. In arXiv:1803.05567

2018 arXiv
[18]

Chris Hokamp and Qun Liu. 2017. Lexically constrained decoding for sequence generation using grid beam search. In ACL

2017
[19]

Sarthak Jain and Byron C. Wallace. 2019. Attention is not explanation. In NAACL

2019
[20]

Samuel L \"a ubli, Rico Sennrich, and Martin Volk. 2018. Has Machine Translation Achieved Human Parity? A Case for Document-level Evaluation . In EMNLP

2018
[21]

Jiwei Li, Will Monroe, and Dan Jurafsky. 2016. Understanding neural networks through representation erasure. In arXiv preprint arXiv:1612.08220

2016 arXiv
[22]

Makoto Morishita, Jun Suzuki, and Masaaki Nagata. 2017. Ntt neural machine translation systems at wat 2017. In WAT

2017
[23]

Pramod Kaushik Mudrakarta, Ankur Taly, Mukund Sundararajan, and Kedar Dhamdhere. 2018. Did the model understand the question? In ACL

2018
[24]

Matt Post and David Vilar. 2018. Fast lexically constrained decoding with dynamic beam allocation for neural machine translation. In NAACL

2018
[25]

Robert Schwarzenberg, David Harbecke, Vivien Macketanz, Eleftherios Avramidis, and Sebastian M \"o ller. 2019. Train, sort, explain: Learning to diagnose translation models. In NAACL

2019
[26]

Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. 2017. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV

2017
[27]

Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In ACL

2016
[28]

Xing Shi, Inkit Padhi, and Kevin Knight. 2016. Does string-based neural mt learn source syntax? In EMNLP

2016
[29]

Emma Strubell, Patrick Verga, Daniel Andor, David Weiss, and Andrew McCallum. 2018. Linguistically-Informed Self-Attention for Semantic Role Labeling . In EMNLP

2018
[30]

Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. In ICML

2017
[31]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NeurIPS

2017
[32]

Wong, Lidia S

Baosong Yang, Longyue Wang, Derek F. Wong, Lidia S. Chao, and Zhaopeng Tu. 2019. Assessing the ability of self-attention networks to learn word order. In ACL

2019
[33]

Wei Emma Zhang, Quan Z Sheng, and Ahoud Abdulrahmn F Alhazmi. 2019. Generating textual adversarial examples for deep learning models: A survey. In arXiv preprint arXiv:1901.06796

2019 arXiv
[34]

Luisa M Zintgraf, Taco S Cohen, Tameem Adel, and Max Welling. 2017. Visualizing deep neural network decisions: Prediction difference analysis. In ICLR

2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

URL: " 'urlintro :=

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

David Alvarez-Melis and Tommi Jaakkola. 2017. A causal framework for explaining the predictions of black-box sequence-to-sequence models. In EMNLP

work page 2017

[4] [4]

Leila Arras, Franziska Horn, Gr \'e goire Montavon, Klaus-Robert M \"u ller, and Wojciech Samek. 2016. Explaining predictions of non-linear classifiers in nlp. In Proceedings of the 1st Workshop on Representation Learning for NLP

work page 2016

[5] [5]

Alexei Baevski and Michael Auli. 2019. Adaptive input representations for neural language modeling. In ICLR

work page 2019

[6] [6]

Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. In ICLR

work page 2014

[7] [7]

Anthony Bau, Yonatan Belinkov, Hassan Sajjad, Nadir Durrani, Fahim Dalvi, and James Glass. 2019. Identifying and controlling important neurons in neural machine translation. In ICLR

work page 2019

[8] [8]

Yonatan Belinkov, Nadir Durrani, Fahim Dalvi, Hassan Sajjad, and James Glass. 2017. What do neural machine translation models learn about morphology? In ACL

work page 2017