Pith. sign in

REVIEW 4 major objections 4 minor 33 references

Main Predicate and Their Arguments as Explanation Signals For Intent Classification

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that the main predicate of an utterance, together with its dependency-grammar arguments, provides reliable word-level explanations for intent classification, and that training models to focus on these signals improves…

desk verdict First real intent-classification explainability benchmark, but the paper's own motivating example contradicts its algorithm's output—that's a load-bearing flaw the authors need to fix. read the letter →

arxiv 2502.01270 v1 pith:MASX2OXJ submitted 2025-02-03 cs.CL

Sameer Pimparkhede , Pushpak Bhattacharyya This is my paper · ORCID
classification cs.CL
keywords intentclassificationexplainabilitydependencyparsingmainpredicateargumentstructuresaliencymapsattributionpriorssilver-standardannotations
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 sets out to show that a cheap linguistic signal—the main predicate of an utterance and the words it governs through specific dependency relations—can replace costly human annotation of explanations for intent classification. The authors build an automatic annotation pipeline, apply it to the ATIS and SNIPS datasets, and contribute about 21,000 utterances with word-level explanation signals, forming a benchmark for explainability. They report that models trained with an attribution prior that rewards focusing on these signals produce saliency maps that match the silver labels better, improving plausibility Token F1 by 3–4 percent on both datasets. The paper also finds that the most accurate classifiers are not the most explainable ones, and that the joint-training signal helps classes with few training samples.

What carries the argument

The central object is the 'main predicate plus its arguments,' extracted from a dependency parse tree. The algorithm selects the parse root as the main predicate, marks its children with the relations direct object, nominal subject, open clausal complement, or oblique (preferring common-noun obliques when several appear), descends one level for compound children that are not proper nouns, and removes stopwords. This yields a binary word-level explanation signal for each token. The same dependency-grounded signal is then used as the target in an auxiliary attribution-prior loss: the model's integrated-gradient attribution scores are averaged over classes, and the squared difference between those scores and the silver signal is added to the cross-entropy loss.

What would settle it

Run the extraction algorithm on a fresh random sample of 1,000 utterances from an intent dataset not used in the paper, have three trained annotators mark intent-relevant words, and compare the agreement with the algorithm's output. If agreement is near chance on that held-out data—or if the joint-training plausibility gain disappears when a different intent dataset replaces ATIS/SNIPS—the central claim would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that the main predicate (usually the main verb) and selected dependency-grammar arguments—direct object, nominal subject, open clausal complement, and certain oblique nominals—carry the information needed to identify intent-relevant words. The paper formalizes this as a tree-traversal algorithm over the dependency parse: start at the root predicate, mark its chosen children, descend one level for compound nouns that are not proper nouns, and strip stopwords. Three expert evaluators reconstructed the correct intent from these extracted segments with 96.3 percent accuracy on ATIS and 97.8 percent on SNIPS (Fleiss' kappa 0.67 and 0.74), which the paper takes as evidence that the silver annotations are high quality. Guiding models to focus on these segments during training increases plausibility Token F1 by 3–4 percent on both datasets while also improving accuracy for low-resource classes.

Load-bearing premise

The load-bearing premise is that the main predicate plus a few dependency children reliably identifies the words that carry intent in virtually every utterance, even though the paper verifies this on only 200 samples per dataset with three judges each.

Editorial extensions

If this is right

  • A reusable word-level explainability benchmark now exists for intent classification, enabling quantitative evaluation of plausibility and faithfulness rather than qualitative case studies.
  • Training small models such as CNN, LSTM, BERT, RoBERTa, and GPT-2 with the attribution prior raises plausibility Token F1 by 3–4 percent on ATIS and SNIPS.
  • Accuracy improves most for rare intent classes such as atis-meal and atis-flight-no, suggesting the signal helps under data scarcity.
  • High classification accuracy does not imply good explanation quality; the two must be measured separately.
  • The annotation method is automatic and requires no extra human labeling, so new utterances can be annotated on the fly.

Reading between the lines

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

  • The same main-predicate heuristic could be tested as a low-cost rationale generator for other single-intent utterance tasks, though predicate-argument structure may be less clean outside request-style dialogue.
  • Because human evaluation covers only 200 samples per dataset, scaling that check to a larger random sample would make the benchmark's quality claims much stronger.
  • The 3–4 percent plausibility gain is measured against the paper's own silver labels; whether end users of a chatbot perceive the improved explanations as more trustworthy remains an open question.
  • The direct object and oblique markers often align with slot boundaries, so the method could plausibly be folded into joint intent-and-slot models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces a method to automatically produce word-level silver explanation labels for intent classification by marking the main predicate and selected dependency-relation arguments (direct object, nsubj, xcomp, obl, and compound children) of each utterance. Applying this to ATIS and SNIPS yields a claimed 21k-instance explainability benchmark. The authors evaluate CNN, LSTM, BERT, RoBERTa, and GPT-2 with LIME and integrated-gradient attributions, using plausibility (Token F1, IOU F1) and faithfulness (comprehensiveness, sufficiency) metrics. They report that joint training with an attribution-prior loss improves plausibility Token F1 by 3-4%, and they support the silver labels with a human evaluation on 200 samples per dataset, achieving 96-98% agreement with gold intents and Fleiss' kappa of 0.67 (ATIS) and 0.74 (SNIPS).

Significance. If the central claim holds, the dataset is a useful, low-cost resource for evaluating and regularizing intent-classification models, and the method is applicable to other intent corpora. The paper has concrete strengths: the annotation procedure is transparent and reproducible, the human evaluation checks a meaningful property of the silver labels, and the plausibility gains are measured on held-out test splits rather than on training data. However, the token-level correctness of the silver labels is not established, and the paper's own motivating example appears to be mishandled by the annotation algorithm. The reported 3-4% improvement is therefore not yet convincingly tied to the desired reasoning behavior, although the general approach remains plausible and worth further development.

major comments (4)
  1. [§4.2, §7] The motivating ATIS example used in Sections 1 and 7 is not handled by the Section 4.2 algorithm. Under the standard dependency parse of 'At the Charlotte airport, how many different types of aircraft are there for us?', the root is 'are', Step 4 selects 'types' (nsubj) and Step 5 selects 'airport' (common-noun obl), while Step 7 only adds compound children, so 'aircraft' (an nmod of 'types') is not marked. The resulting silver explanation therefore contains 'airport', the exact token that Section 7 identifies as a misleading frequency distractor, and omits 'aircraft', the token that Section 7 calls required. This internal inconsistency weakens the central claim that the algorithm marks exactly the intent-defining words and suggests the reported plausibility gains may partly reflect training on noisy labels.
  2. [§4.3, §6.1] The human evaluation protocol validates sufficiency, not token-level completeness or exclusion. Evaluators were shown only the silver segment and asked to choose an intent, so they could not detect omitted intent-defining tokens or included distracting tokens. The 96-98% agreement therefore does not establish that the silver labels are correct as token-level ground truth, which is how they are used in the plausibility metrics of Tables 2 and 3. Please add a token-level assessment or a systematic error analysis of the heuristic's precision and recall against a smaller, fully annotated sample.
  3. [§6.2, Appendix A.2] The reported 3-4% plausibility improvement lacks variance and selection details. Appendix A.2 says λ was searched over the range 1 to 10^10 and the values with the best results were selected, but it does not state whether this selection was made on the dev set or the test set. If the latter, the improvements in Tables 2 and 3 may be inflated by test-set tuning. Please report means and standard deviations over multiple seeds and specify the model-selection criterion and split.
  4. [§6.2, Tables 2 and 3] The faithfulness results are not uniformly improved by joint training, and this is not discussed. For example, Table 2 shows BERT Joint comprehensiveness on ATIS falling from 0.49 to 0.45, and Table 3 shows BERT Joint comprehensiveness on SNIPS falling from 0.72 to 0.69. The paper should report per-configuration analyses or qualify the claim that joint training improves explainability, rather than presenting the plausibility gain as a general improvement.
minor comments (4)
  1. [Abstract, §1, §3] The dataset size is reported as 21k in the Abstract and Section 3 but as 20k in the contribution list in Section 1; please reconcile these numbers.
  2. [Appendix A.2] The sentence 'For joint attribution training of CNN and LSTM, λ is set to 105 and 104 for the SNIPS dataset' does not specify which model receives which value; please clarify the exact assignments.
  3. [§6.2] The sentence beginning 'We also observe that joint training the performance of models on the faithfulness metric' is incomplete and should be rephrased.
  4. [§6.1] Tables 2 and 3 mix signed values for sufficiency without explaining the sign convention; since Sufficiency is defined as m(x_i)_j - m(r_i)_j, a brief note on why some values are negative would help readers interpret the metric.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported 3-4% plausibility improvement is evaluated against the same silver annotations that the joint loss is trained to reproduce, so the 'improved reasoning' claim largely reduces to fitting the Section 4.2 heuristic.

  1. fitted input called prediction [Section 5.1 (Eq. 3), Section 6.1, Section 6.2]
    "Instead, we use local explanation signals as each sample’s target values. ... Ljoint = L(y, p) + λ Σ (ai − ti)^2 (3) ... Plausibility compares the model’s reasoning with the ground truth explanation. Here, we use a silver annotated explanation as ground truth."

    The joint loss (Eq. 3) sets ti to the silver explanation signals and penalizes integrated-gradient attributions ai that deviate from ti. The plausibility metric (Token F1, Sec. 6.1) is then computed against 'a silver annotated explanation as ground truth' — the same ti values. Consequently, the 3-4% Token F1 gain reported in Sec. 6.2 and the abstract measures how well the model reproduces the Section 4.2 main-predicate-plus-arguments heuristic after being explicitly regularized to do exactly that. The human evaluation (Sec. 4.3) does not break the loop: it gives annotators only the silver segment and asks them to select an intent, which tests whether the segment is sufficient for the label, not whether it contains every intent-defining token and excludes distractors.

full rationale

The silver-annotation algorithm itself is not circular: Section 4.2 defines a dependency-based heuristic (main predicate, direct object, nsubj, xcomp, selected obl, compound children, stopword removal) that is independent of the model outputs, and the 200-sample-per-dataset human study in Section 4.3 provides a genuine, if partial, external check. There are no load-bearing self-citations or imported uniqueness theorems. The circularity is concentrated in the paper's central empirical claim: the joint attribution loss (Eq. 3) trains models to align integrated-gradient attributions with silver explanation signals, and the same silver signals are then used as ground truth for the plausibility Token F1 improvement (Secs. 6.1 and 6.2). Because a model regularized to match ti on training samples is expected to match ti on similar test samples, the 3-4% plausibility gain is largely a measure of fit to the annotation algorithm rather than evidence of more human-like reasoning. The human validation only establishes that the silver segment is often sufficient to identify the intent; it does not establish token-level precision or the absence of misleading words. A separate correctness concern (not itself circularity) reinforces this: on the paper's motivating ATIS example, the Section 4.2 rules would mark 'types' and 'airport' as explanation signals while 'aircraft' (an nmod of 'types') is not collected, so the silver label may contain the very distractor the paper says misleads BERT. Taking the silver labels as ground truth for both training and evaluation is therefore the key circular step, and it affects the headline 3-4% result rather than a peripheral claim.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central method rests on hand-picked linguistic heuristics and tooling assumptions rather than on new theoretical entities. The free parameters are the attribution loss weight, the top-k evaluation threshold, and the categorical choices of dependency relations and POS constraints. No new physical or conceptual entities are introduced.

free parameters (5)
  • Attribution prior weight λ = 10^4 to 10^6 depending on model and dataset (e.g., 10^6 for ATIS CNN/LSTM, 10^5 and 10^4 for SNIPS CNN and LSTM)
    Tuned by extensive search over (1, 10^10) for 'best results'; selection set (dev or test) is not specified, so this is a hand-fitted parameter affecting the joint loss.
  • Top-k words for plausibility evaluation = 5
    Chosen because silver explanations average 3.4-4.2 words; this threshold directly determines Token F1 and IOU F1 scores.
  • Dependency relation set for explanation signals = dobj, nsubj, xcomp, obl
    Hand-picked based on examples in Section 4.2; other relations (e.g., nmod, amod) are excluded with post-hoc reasoning.
  • POS constraints on obl and compound children = obl must be common noun; compound must not be proper noun
    Added to avoid proper-noun noise, motivated by examples where named entities do not determine intent.
  • Stopword list = Python NLTK default stopwords
    Applied to reduce noise; list is not customized to the task, potentially removing meaningful words.
assumptions (4)
  • domain assumption The main verb denotes the action and the direct object indicates the domain of conversation (§1, §4.2).
    This observation motivates the entire annotation algorithm and is stated as a hypothesis rather than proven for all utterances.
  • domain assumption OpenIE and Stanford CoreNLP produce accurate dependency parses for these short, informal utterances (Figure 1, §4.2).
    The algorithm depends on parser output for every utterance; no parser accuracy is reported on the datasets.
  • domain assumption Human evaluators can infer intent correctly from the explanation segment alone (§4.3).
    The quality of silver explanations is judged by asking experts to label intent from the highlighted words, assuming that the segment is sufficient for that task.
  • domain assumption Words outside the selected explanation segment are not necessary for intent classification (implicit in the annotation).
    The algorithm marks only a subset of tokens; tokens like proper-noun obliques and adjectives are excluded with the rationale that they do not affect intent, which is not verified exhaustively.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Main Predicate and Their Arguments as Explanation Signals For Intent Classification." pith.science (2026). https://pith.science/paper/MASX2OXJ

@misc{pith2026250201270,
  author       = {Pith},
  title        = {Pith review of: Main Predicate and Their Arguments as Explanation Signals For Intent Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MASX2OXJ}},
  note         = {Machine review of arXiv:2502.01270}
}
read the original abstract

Intent classification is crucial for conversational agents (chatbots), and deep learning models perform well in this area. However, little research has been done on the explainability of intent classification due to the absence of suitable benchmark data. Human annotation of explanation signals in text samples is time-consuming and costly. However, from inspection of data on intent classification, we see that, more often than not, the main verb denotes the action, and the direct object indicates the domain of conversation, serving as explanation signals for intent. This observation enables us to hypothesize that the main predicate in the text utterances, along with the arguments of the main predicate, can serve as explanation signals. Leveraging this, we introduce a new technique to automatically augment text samples from intent classification datasets with word-level explanations. We mark main predicates (primarily verbs) and their arguments (dependency relations) as explanation signals in benchmark intent classification datasets ATIS and SNIPS, creating a unique 21k-instance dataset for explainability. Further, we experiment with deep learning and language models. We observe that models that work well for classification do not perform well in explainability metrics like plausibility and faithfulness. We also observe that guiding models to focus on explanation signals from our dataset during training improves the plausibility Token F1 score by 3-4%, improving the model's reasoning.

Figures

Figures reproduced from arXiv: 2502.01270 by the authors.

Figure 1
Figure 1. Dependancy parse tree of text sample from SNIPS dataset having a label as BookRestaurants. Highlighted [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 10 canonical work pages

  1. [1]

    Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.263 A diagnostic study of explainability techniques for text classification . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3256--3274, Online. Association for Computati...

    doi:10.18653/v1/2020.emnlp-main.263 2020

  2. [2]

    I \ n igo Casanueva, Tadas Tem c inas, Daniela Gerz, Matthew Henderson, and Ivan Vuli \'c . 2020. https://doi.org/10.18653/v1/2020.nlp4convai-1.5 Efficient intent detection with dual sentence encoders . In Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, pages 38--45, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.nlp4convai-1.5 2020

  3. [3]

    Qian Chen, Zhu Zhuo, and Wen Wang. 2019. Bert for joint intent classification and slot filling. arXiv preprint arXiv:1902.10909

    arXiv 2019

  4. [4]

    Alice Coucke, Alaa Saade, Adrien Ball, Th \'e odore Bluche, Alexandre Caulier, David Leroy, Cl \'e ment Doumouro, Thibault Gisselbrecht, Francesco Caltagirone, Thibaut Lavril, et al. 2018. Snips voice platform: an embedded spoken language understanding system for private-by-design voice interfaces. arXiv preprint arXiv:1805.10190

    arXiv 2018

  5. [5]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long a...

    doi:10.18653/v1/n19-1423 2019

  6. [6]

    Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, and Byron C Wallace. 2019. Eraser: A benchmark to evaluate rationalized nlp models. arXiv preprint arXiv:1911.03429

    arXiv 2019

  7. [7]

    Yanai Elazar, Shauli Ravfogel, Alon Jacovi, and Yoav Goldberg. 2021. https://doi.org/10.1162/tacl_a_00359 Amnesic probing: Behavioral explanation with amnesic counterfactuals . Transactions of the Association for Computational Linguistics, 9:160--175

    doi:10.1162/tacl_a_00359 2021

  8. [8]

    Kalpa Gunaratna, Vijay Srinivasan, Akhila Yerukola, and Hongxia Jin. 2022. Explainable slot type attentions to improve joint intent detection and slot filling. arXiv preprint arXiv:2210.10227

    work page Pith review arXiv 2022

Show all 33 references
  1. [9]

    Shirley Anugrah Hayati, Dongyeop Kang, and Lyle Ungar. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.510 Does BERT learn as humans perceive? understanding linguistic styles through lexica . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Proc...

    2021 doi

  2. [10]

    Hemphill, John J

    Charles T. Hemphill, John J. Godfrey, and George R. Doddington. 1990. https://aclanthology.org/H90-1021 The ATIS spoken language systems pilot corpus . In Speech and Natural Language: Proceedings of a Workshop Held at Hidden Valley, P ennsylvania, June 24-27,1990

    1990

  3. [11]

    Sepp Hochreiter and J \"u rgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735--1780

    1997

  4. [12]

    Sahil Jayaram and Emily Allaway. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.450 Human rationales as attribution priors for explainable stance detection . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5540--5554, Online ...

    2021 doi

  5. [13]

    Ratnesh Joshi, Arindam Chatterjee, and Asif Ekbal. 2021. Towards explainable dialogue system: Explaining intent classification using saliency techniques. In Proceedings of the 18th International Conference on Natural Language Processing (ICON), pages 120--127

    2021

  6. [14]

    Yoon Kim. 2014. Convolutional neural networks for sentence classification. arXiv preprint arXiv:1408.5882

    2014 arXiv

  7. [15]

    J Richard Landis and Gary G Koch. 1977. The measurement of observer agreement for categorical data. biometrics, pages 159--174

    1977

  8. [16]

    Shiyang Li, Jianshu Chen, Yelong Shen, Zhiyu Chen, Xinlu Zhang, Zekun Li, Hong Wang, Jing Qian, Baolin Peng, Yi Mao, et al. 2022. Explanations from large language models make small reasoners better. arXiv preprint arXiv:2210.06726

    2022 arXiv

  9. [17]

    Frederick Liu and Besim Avci. 2019. Incorporating priors with feature attribution on text classification. arXiv preprint arXiv:1906.08286

    2019 arXiv

  10. [18]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

    2019 arXiv

  11. [19]

    Binny Mathew, Punyajoy Saha, Seid Muhie Yimam, Chris Biemann, Pawan Goyal, and Animesh Mukherjee. 2021. Hatexplain: A benchmark dataset for explainable hate speech detection. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 14867--14875

    2021

  12. [20]

    Gr \'e goire Montavon, Alexander Binder, Sebastian Lapuschkin, Wojciech Samek, and Klaus-Robert M \"u ller. 2019. Layer-wise relevance propagation: an overview. Explainable AI: interpreting, explaining and visualizing deep learning, pages 193--209

    2019

  13. [21]

    Libo Qin, Tailu Liu, Wanxiang Che, Bingbing Kang, Sendong Zhao, and Ting Liu. 2021. A co-interactive transformer for joint slot filling and intent detection. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8193--8197. IEEE

    2021

  14. [22]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9

    2019

  15. [23]

    Christian Raymond and Giuseppe Riccardi. 2007. Generative and discriminative algorithms for spoken language understanding. In Interspeech 2007-8th Annual Conference of the International Speech Communication Association

    2007

  16. [24]

    why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. " why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135--1144

    2016

  17. [25]

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2018. Anchors: High-precision model-agnostic explanations. In Proceedings of the AAAI conference on artificial intelligence, volume 32

    2018

  18. [26]

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. In International conference on machine learning, pages 3319--3328. PMLR

    2017

  19. [27]

    Yequan Wang, Minlie Huang, Xiaoyan Zhu, and Li Zhao. 2016. https://doi.org/10.18653/v1/D16-1058 Attention-based LSTM for aspect-level sentiment classification . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 606--615, Austin, T...

    2016 doi

  20. [28]

    Zhengxuan Wu and Desmond C Ong. 2021. On explaining your explanations of bert: An empirical study with sequence classification. arXiv preprint arXiv:2101.00196

    2021 arXiv

  21. [29]

    Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. 2015. https://proceedings.mlr.press/v37/xuc15.html Show, attend and tell: Neural image caption generation with visual attention . In Proceedings of the 32nd Int...

    2015

  22. [30]

    Xi Ye and Greg Durrett. 2022. The unreliability of explanations in few-shot prompting for textual reasoning. Advances in neural information processing systems, 35:30378--30392

    2022

  23. [31]

    Ruiqi Zhong, Steven Shao, and Kathleen McKeown. 2019. Fine-grained sentiment analysis with faithful attention. arXiv preprint arXiv:1908.06870

    2019 arXiv

  24. [32]

    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...

  25. [33]

    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 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.