Pith. sign in

REVIEW 4 major objections 6 minor 20 references

Model-free Speculative Decoding for Transformer-based ASR with Token Map Drafting

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Token Map Drafting replaces the draft model with an n-gram token map, claiming 1.27x–1.37x CPU decoding speedups for Whisper without accuracy loss.

desk verdict Model-free n-gram drafting for ASR spec decoding is a nice idea with plausible CPU speedups, but the missing WER numbers leave the accuracy-preservation claim unproven. read the letter →

arxiv 2507.21522 v1 pith:MQPXXJFE submitted 2025-07-29 cs.CL cs.SDeess.AS

Tuan Vu Ho , Hiroaki Kokubo , Masaaki Yamamoto , Yohei Kawaguchi This is my paper · ORCID
classification cs.CLcs.SDeess.AS
keywords speculativedecodingautomaticspeechrecognitionWhispern-gramtokenmapmodel-freedraftingCPUinferenceon-deviceASRlow-perplexitydomains
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 proposes Token Map Drafting, a speculative decoding method for transformer ASR that replaces the usual small draft model with a precomputed n-gram token map built from domain-specific transcripts. The aim is to accelerate autoregressive decoding of Whisper-family models on CPU-only devices, where running a second model would add overhead. The paper claims 1.27x and 1.37x decoding speedups on the CI-AVSR dataset and an internal maintenance-task dataset, with no reported degradation in transcription accuracy, plus a 10% absolute speed improvement over a distilled-draft-model baseline. The central contribution is showing that, for structured low-perplexity speech, domain statistics can substitute for a learned drafter.

What carries the argument

The central object is the n-gram token map: a dictionary whose keys are token n-grams (length 1 to N) and whose values are ranked lists of candidate token sequences that followed those n-grams in the domain training transcripts. It replaces the learned draft model of standard speculative decoding; decoding runs by matching the current token context to a key, proposing the stored continuation as a draft, verifying all proposed tokens with one parallel pass of the main decoder, and keeping tokens until the first mismatch. The matching and verification loop, together with the pruning rule derived from the paper's timing analysis (only propose when candidate sequences are long enough to beat autoregressive cost), carries the argument.

What would settle it

Measure word error rate on both datasets with Token Map Drafting versus unmodified Whisper decoding on the same audio; if the WER rises, the central claim of no accuracy loss falls. The paper provides speed and acceptance metrics but no accuracy table, so this measurement is the decisive check.

Watch

Extended reading notes

Core claim

Token Map Drafting is a model-free speculative decoding algorithm for Whisper-family ASR. A token map is constructed by tokenizing domain training transcripts, extracting n-grams of lengths 1 through N as keys, recording the token sequences that follow each n-gram as candidate drafts, and pruning candidates by length and frequency. At decode time, the current decoder state's last tokens are matched against map keys; matching candidate continuations are proposed, verified in parallel by the main model, and accepted or truncated back to the first unmatched token. The paper reports 1.27x and 1.37x decoding speedups on CI-AVSR and an internal maintenance-command dataset, and claims these come without degrading transcription accuracy and with a 10% absolute speed advantage over a distilled-draft-model baseline on CPU.

Load-bearing premise

The central premise is that accepting draft tokens whose conditional probability clears a fixed threshold yields exactly the tokens the main model would have produced; the paper does not report any accuracy metric, so the claim that accuracy is preserved depends on that output-equivalence assumption.

Editorial extensions

If this is right

  • On CI-AVSR, Token Map Drafting speeds Whisper-large-v3 decoding by 1.27x compared with autoregressive decoding, with the paper reporting no transcription accuracy degradation.
  • On the internal maintenance-task dataset, the speedup reaches 1.37x with Whisper-large-v3, and 1.35x to 1.36x with Whisper-medium and Whisper-small, so the gain is not tied to the largest model.
  • Compared with a distilled Whisper draft model on CPU, the token map gives roughly a 10 percentage point speedup advantage on the internal dataset, despite a shorter average acceptance length, showing that the draft model's overhead was a major bottleneck.
  • Because no draft model needs to be trained or loaded, the method is directly deployable on CPU-only and memory-constrained devices for domain-specific ASR.
  • The method is designed for structured, low-perplexity domains such as in-car commands or measurement reports; its benefit shrinks on diverse, spontaneous speech, where the reported CI-AVSR acceptance rate is lower.

Reading between the lines

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

  • Editorial extension: because the token map is built from domain training transcripts, the same drafting idea should transfer to other autoregressive sequence models in closed-vocabulary, structured domains, but only if the tokenizer and domain text statistics are stable enough that n-gram continuations are reliable.
  • Editorial extension: the headline speedups are averages over utterances; in a long tail of utterances with no matching n-gram, the method falls back to autoregressive decoding, so tail latency could be much worse than the mean suggests.
  • Editorial extension: the optimal n-gram length and pruning threshold are likely domain-dependent; the paper measures a peak at N=3 on one internal dataset, and a different domain perplexity could shift that optimum.
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 / 6 minor

Summary. The paper proposes Token Map Drafting, a model-free speculative decoding method for transformer-based ASR (Whisper). Instead of a separate draft model, it builds a precomputed n-gram token map from domain-specific transcriptions and uses it to propose candidate token sequences. The main model verifies these candidates by thresholding their conditional probabilities, and accepted tokens are kept while mismatches trigger autoregressive correction. Experiments on the CI-AVSR dataset and an internal maintenance-task dataset report decoding speedups of 1.27x and 1.37x over autoregressive decoding, and a comparison to a Distil-Whisper-based speculative decoding baseline. The paper claims these speedups are obtained 'without sacrificing transcription accuracy', but no word error rate or any accuracy metric is reported anywhere.

Significance. If the accuracy-preservation claim could be substantiated, the method would be a practical contribution for CPU-based ASR in structured, low-perplexity domains, since it removes the need for an auxiliary draft model and exploits domain-specific regularities. The core idea is simple and plausible, and the reported speedup numbers, taken at face value, are modest but potentially useful. However, the significance is almost entirely contingent on empirical validation that is currently missing: there are no WER/CER numbers, the acceptance threshold is not disclosed, and the hyperparameters are selected and evaluated on the same internal dataset. Without these components, the paper does not establish its central value proposition.

major comments (4)
  1. [Abstract; Table I; Section IV-B] The claim 'without sacrificing transcription accuracy' is unsupported because no accuracy metric is reported. Table I reports only speedup (S), acceptance rate (A_r), and average acceptance length (A_l); no WER, CER, or any recognition quality metric appears for the baseline or the proposed method. The abstract and conclusion assert accuracy preservation, but the evaluation metrics in Section IV-B do not include accuracy. The paper must add a direct comparison of transcription accuracy (e.g., WER/CER) for both the baseline and the proposed method before this claim can be accepted.
  2. [Section II, Step 3; Section III] The verification rule described in Section II (and reflected in Figure 2) accepts a draft token if its conditional probability exceeds a predefined threshold, otherwise it generates a replacement token from the main model. This is not the distribution-preserving rejection scheme of Leviathan et al. [12]. With greedy decoding, which is the typical mode for CTranslate2 Whisper, the main model's output at each position is its argmax; accepting a non-argmax token whose probability merely exceeds an unreported threshold changes the final transcription. The threshold value is never given, so the reader cannot tell whether the method is configured to preserve outputs or to trade accuracy for speed. The authors should either specify the threshold and demonstrate empirically that WER is unchanged, or change the acceptance rule to one that guarantees output equivalence.
  3. [Section III-B, Figure 4, Table I] The hyperparameter selection is circular with respect to the evaluation. Section III-B states that the optimal n-gram length N is chosen by measuring speedup on 'an internal dataset' (Figure 4), and the pruning thresholds shown in Figure 3 appear to be derived from similar speedup measurements. The final speedup for the proposed method on the internal dataset (Table I) is then reported on the same data used to select N and the thresholds. This means the reported 1.37x speedup is partially optimized on the evaluation set. The authors should use a held-out validation split or cross-validation for hyperparameter selection and report speedups on a separate test set.
  4. [Algorithm 1; Section III; Figure 2] The runtime decoding algorithm is not fully specified. Algorithm 1 only describes token-map construction; the decoding-time steps of sequence matching, validation, candidate merging, and acceptance are only described informally in the text and Figure 2. In particular, the acceptance threshold, the procedure for 'iteratively merge candidates with their nearest matches', and the condition under which decoder state is truncated are not formalized. Without a precise algorithmic statement, the method is not reproducible and the behavior of the verification step cannot be audited. A complete pseudo-code for the decoding loop should be added.
minor comments (6)
  1. [Abstract] There is a typographical artifact in the first line: 'we proposeT oken Map Drafting' should be 'we propose Token Map Drafting'.
  2. [Abstract; Conclusion] The phrase '10% absolute improvement in decoding speed' is misleading. Table I shows speedups of 1.27 for the baseline and 1.37 for the proposed method, a difference of 0.10 in the speedup factor. This is a 0.10x speedup difference, or about 7.9% relative improvement, not '10% absolute improvement' in decoding speed.
  3. [Section III-B] The sentence 'Beyond 3 candidates, SD show less effective' contains a subject-verb agreement error and should be rewritten, e.g., 'Beyond 3 candidates, SD is less effective'.
  4. [Table I] The table header formatting is inconsistent ('CI-A VSR' versus 'Internal Dataset'), and the column alignment is off in the rendered version. Please format the table cleanly and ensure the dataset names match the text.
  5. [Conclusion] The sentence 'While our experimental results suggest that the proposed method maintains transcription accuracy' is not supported by any reported accuracy result. This sentence should be revised or removed until actual WER/CER measurements are reported.
  6. [Section IV-A] The baseline configuration should be described more carefully. The paper cites Distil-Whisper [15] as the basis of the 'Distill-spec' baseline, but uses Whisper-large-v3-turbo as the draft model. It is not clear whether this exact setup is prescribed by [15] or is a modified configuration; please clarify the baseline hyperparameters and decoding settings.

Circularity Check

2 steps flagged · score 6.0 of 10

The internal-dataset speedup is partially circular because the n-gram length and pruning thresholds were tuned on that same dataset; the accuracy-preservation claim is asserted without WER support.

  1. fitted input called prediction [Section III-B, Figure 4, and Table I]
    "To determine the optimal n-gram length N, we measure speedup rates across different values of N using Whisper-small on an internal dataset. Figure 4 shows that speedup peaks at N=3, which we adopt for subsequent experiments."

    The paper later reports the internal-dataset speedup of 1.37x (Table I) as a headline result, but N=3 was selected by measuring speedup on that same internal dataset. The hyperparameter is therefore fit to the evaluation set, and the reported speedup is the value of the objective used for the selection rather than an independent held-out prediction. The CI-AVSR result is less affected because N was chosen on a different dataset, but the central internal-dataset claim is not independent of the tuning procedure.

  2. fitted input called prediction [Section III-B, Figure 3, and Algorithm 1 line 10]
    "Since SD does not always outperform standard autoregressive decoding in every scenario, we empirically analyze inference time under varying candidate numbers and sequence lengths. Figure 3 shows that SD achieves speedup only when candidate sequence lengths exceed certain thresholds (e.g., 9 tokens for 2 candidates, 16 for 3 candidates). Beyond 3 candidates, SD show less effective. Based on these findings, we iteratively merge candidates with their nearest matches until the condition is met."

    The number of candidate sequences and the minimum sequence-length thresholds are chosen from the same empirical decoding-time analysis, and Algorithm 1 then prunes the token map using 'the condition in Figure 3'. Table I is produced with these pruned settings on the internal dataset, so the reported speedup is conditional on thresholds fit to the same evaluation data. No separate validation set is reported, making the internal-dataset speedup a fitted input rather than a neutral prediction.

full rationale

The paper's central contribution is an empirical speedup claim, and the derivation chain is otherwise self-contained: Token Map Drafting constructs an n-gram token map from domain transcriptions and uses it as a draft source, with no separate trained draft model. The main circularity concern is test-set tuning: the n-gram length N and the pruning thresholds/candidate counts are selected by measuring speedup on the internal dataset, and the same internal dataset is then used for the headline 1.37x speedup and the 10% improvement over the Distill-spec baseline. This means those numbers are not independent evaluations of a fixed method; they are partly the result of optimizing on the evaluation set. The CI-AVSR result (1.27x) provides some external validation because N was chosen on the internal dataset rather than on CI-AVSR, but the paper does not describe a held-out split for the internal dataset, so the main numeric claim remains contaminated. Separately, the abstract's assertion of 'without sacrificing transcription accuracy' is not supported by any WER or accuracy metric in Table I or elsewhere. The verification step accepts any token whose conditional probability exceeds an unreported threshold, which can alter the greedy decoding output, so the accuracy-preservation claim is a correctness risk. I do not count this as a circular step because it is an unsupported assertion rather than a derivation that reduces to its inputs. There is no load-bearing self-citation chain, and no uniqueness theorem or external authority is invoked to force the method. Overall, the partial circularity from tuning on the reported dataset justifies a score of 6 rather than a higher score, because the method itself is not definitionally equivalent to its inputs and the CI-AVSR benchmark is external, but the internal-dataset speedup is not an independent confirmation.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central speedup claim depends on hyperparameters chosen on the evaluation set (N=3, pruning thresholds) and on the domain assumption that transcripts are low-perplexity. The accuracy claim depends on an unproven threshold-acceptance rule. No code or data is released.

free parameters (3)
  • N (maximum n-gram length) = 3
    Chosen as the value that maximizes speedup on the internal dataset (Figure 4); reported speedups use this tuned value, so the headline result is optimized on the evaluation set.
  • Candidate count K and sequence length thresholds = not reported
    Figure 3 is used to decide how many candidates to keep and which sequence lengths are worth speculating; the merge and prune rule in Algorithm 1 line 10 is not specified quantitatively.
  • Acceptance probability threshold = not reported
    Section II states draft tokens are accepted when their probability is above a predefined threshold, but the threshold value is never given; it controls the trade-off between speed and output fidelity.
assumptions (3)
  • domain assumption Target ASR transcripts are structured or low-perplexity so that n-gram continuations from a training map match decoder context often enough to speed up decoding.
    Used in Sections I and III.A to justify the method; if the test domain is diverse, acceptance rate drops (e.g., 38.6% on CI-AVSR) and gains shrink.
  • domain assumption The token map built from domain training transcriptions generalizes to test transcriptions without leakage.
    Algorithm 1 constructs the map from training transcriptions D; the paper does not specify dataset splits or whether the internal test set overlaps in vocabulary with the map source.
  • ad hoc to paper Threshold-based acceptance of draft tokens preserves the main model's output distribution or transcription accuracy.
    Section II Step 3 describes accepting tokens with probability above a threshold rather than matching the main model's greedy choice; no proof or WER measurement is supplied.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Model-free Speculative Decoding for Transformer-based ASR with Token Map Drafting." pith.science (2026). https://pith.science/paper/MQPXXJFE

@misc{pith2026250721522,
  author       = {Pith},
  title        = {Pith review of: Model-free Speculative Decoding for Transformer-based ASR with Token Map Drafting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MQPXXJFE}},
  note         = {Machine review of arXiv:2507.21522}
}
abstract

End-to-end automatic speech recognition (ASR) systems based on transformer architectures, such as Whisper, offer high transcription accuracy and robustness. However, their autoregressive decoding is computationally expensive, hence limiting deployment on CPU-based and resource-constrained devices. Speculative decoding (SD) mitigates this issue by using a smaller draft model to propose candidate tokens, which are then verified by the main model. However, this approach is impractical for devices lacking hardware accelerators like GPUs. To address this, we propose \emph{Token Map Drafting}, a model-free SD technique that eliminates the need for a separate draft model. Instead, we leverage a precomputed n-gram token map derived from domain-specific training data, enabling efficient speculative decoding with minimal overhead. Our method significantly accelerates ASR inference in structured, low-perplexity domains without sacrificing transcription accuracy. Experimental results demonstrate decoding speed-ups of $1.27\times$ on the CI-AVSR dataset and $1.37\times$ on our internal dataset without degrading recognition accuracy. Additionally, our approach achieves a $10\%$ absolute improvement in decoding speed over the Distill-spec baseline running on CPU, highlighting its effectiveness for on-device ASR applications.

Figures

Figures reproduced from arXiv: 2507.21522 by the authors.

Figure 1
Figure 1. Overview of speculative decoding. Divergent tokens is highlight in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed speculative decoding with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Speedup analysis of processing times with varying n-gram length [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 11 canonical work pages

  1. [12]

    Fast Inference from Transformers via Speculative Decoding,

    Y . Leviathan, M. Kalman, and Y . Matias, “Fast Inference from Transformers via Speculative Decoding,” inProceedings of the 40th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett, Eds., vol. 202. PMLR, 23–29 Jul 2023, pp. 19 274–19 286. [O...

    work page 2023

  2. [1]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  3. [2]

    Robust speech recognition via large-scale weak supervi- sion,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervi- sion,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

    2023

  4. [3]

    Conformer: Convolution- augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution- augmented transformer for speech recognition,” inProcs. of INTER- SPEECH, 2020

    work page 2020

  5. [4]

    SpeechT5: Unified- Modal Encoder-Decoder Pre-Training for Spoken Language Processing,

    J. Ao, R. Wang, L. Zhou, C. Wang, S. Ren, Y . Wu, S. Liu, T. Ko, Q. Li, Y . Zhang, Z. Wei, Y . Qian, J. Li, and F. Wei, “SpeechT5: Unified- Modal Encoder-Decoder Pre-Training for Spoken Language Processing,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, 2022, pp. 5723–5738

    work page 2022

  6. [5]

    FlashAttention: Fast and memory-efficient exact attention with IO-awareness,

    T. Dao, D. Y . Fu, S. Ermon, A. Rudra, and C. R ´e, “FlashAttention: Fast and memory-efficient exact attention with IO-awareness,” inAdvances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  7. [6]

    FlashAttention-2: Faster attention with better parallelism and work partitioning,

    T. Dao, “FlashAttention-2: Faster attention with better parallelism and work partitioning,” inInternational Conference on Learning Represen- tations (ICLR), 2024

    2024

  8. [7]

    Folding Attention: Memory and Power Optimization for On-Device Transformer-based Streaming Speech Recognition,

    Y . Li, L. Lai, Y . Shangguan, F. N. Iandola, Z. Ni, E. Chang, Y . Shi, and V . Chandra, “Folding Attention: Memory and Power Optimization for On-Device Transformer-based Streaming Speech Recognition,” in Proc. of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 901–11 905

    work page 2024

Show all 20 references
  1. [8]

    Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale,

    T. Dettmers, M. Lewis, Y . Belkada, and L. Zettlemoyer, “Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale,”Advances in neural information processing systems, vol. 35, pp. 30 318–30 332, 2022

    2022

  2. [9]

    2-bit Conformer quantization for automatic speech recognition,

    O. Rybakov, P. Meadowlark, S. Ding, D. Qiu, J. Li, D. Rim, and Y . He, “2-bit Conformer quantization for automatic speech recognition,” in Proc. of INTERSPEECH, 2023

    2023

  3. [10]

    Minilm: Deep self-attention distillation for task-agnostic compression of pre- trained transformers,

    W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, and M. Zhou, “Minilm: Deep self-attention distillation for task-agnostic compression of pre- trained transformers,”Advances in neural information processing sys- tems, vol. 33, pp. 5776–5788, 2020

    2020

  4. [11]

    Tinybert: Distilling bert for natural language understanding,

    X. Jiao, Y . Yin, L. Shang, X. Jiang, X. Chen, L. Li, F. Wang, and Q. Liu, “Tinybert: Distilling bert for natural language understanding,” inFindings of the Association for Computational Linguistics: EMNLP

  5. [13]

    Acceler- ating large language model decoding with speculative sampling,

    C. Chen, S. Borgeaud, G. Irving, J.-B. Lespiau, and L. Sifre, “Acceler- ating large language model decoding with speculative sampling,”arXiv preprint arXiv:2302.01318, 2023

    2023 arXiv

  6. [14]

    Online speculative decoding,

    X. Liu, L. Hu, P. Bailis, A. Cheung, Z. Deng, I. Stoica, and H. Zhang, “Online speculative decoding,” inProc. of the 41st International Con- ference on Machine Learning. JMLR.org, 2024

    2024

  7. [15]

    Distil-whisper: Robust knowledge distillation via large-scale pseudo labelling,

    S. Gandhi, P. von Platen, and A. M. Rush, “Distil-whisper: Robust knowledge distillation via large-scale pseudo labelling,”arXiv preprint arXiv:2311.00430, 2023

    2023 arXiv

  8. [16]

    Whisper in Medusa’s Ear: Multi-head Efficient Decoding for Transformer-based ASR,

    Y . Segal-Feldman, A. Shamsian, A. Navon, G. Hetz, and J. Keshet, “Whisper in Medusa’s Ear: Multi-head Efficient Decoding for Transformer-based ASR,”arXiv preprint arXiv:2409.15869, 2024

    2024 arXiv

  9. [17]

    LibriSpeech: an ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “LibriSpeech: an ASR corpus based on public domain audio books,” in2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  10. [18]

    V oxPopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation,

    C. Wang, M. Riviere, A. Lee, A. Wu, C. Talnikar, D. Haziza, M. Williamson, J. Pino, and E. Dupoux, “V oxPopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation,”Annual Meeting of the Association for Computationa...

    2021

  11. [19]

    CI-A VSR: A Cantonese Audio-Visual Speech Dataset for In- car Command Recognition,

    W. Dai, S. Cahyawijaya, T. Yu, E. J. Barezi, P. Xu, C. T. S. Yiu, R. Frieske, H. Lovenia, G. I. Winata, Q. Chen, X. Ma, B. E. Shi, and P. Fung, “CI-A VSR: A Cantonese Audio-Visual Speech Dataset for In- car Command Recognition,”ArXiv, vol. abs/2201.03804, 2022

    2022 arXiv

  12. [2020]

    Association for Computational Linguistics, 2020

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.