Pith. sign in

REVIEW 4 major objections 6 minor 38 references

Handling Korean Out-of-Vocabulary Words with Phoneme Representation Learning

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 KOPL shows that adding phoneme representations to Korean word embeddings improves out-of-vocabulary handling by an average of 1.9% across five NLP tasks.

desk verdict KOPL is a believable, useful method paper that adds IPA phoneme representations to Korean OOV handling; the phonemizer is unvalidated, but the core idea holds up and the experiments are solid enough to warrant peer review. read the letter →

arxiv 2507.04018 v1 pith:GOX5KE6H submitted 2025-07-05 cs.CL

Nayeon Kim , Eojin Jeon , Jun-Hyung Park , SangKeun Lee This is my paper
classification cs.CL
keywords Out-of-VocabularyPhonemerepresentationKoreanNLPWordembeddingsContrastivelearningHangeulOOVwordsMultimodal
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 claims that Korean out-of-vocabulary (OOV) words can be handled better by learning phoneme representations alongside word representations and fusing them. Because Hangul is a phonemic script, the spelling of a Korean word carries information about its pronunciation, and pronunciation can link an unseen word to known words that sound alike. The proposed framework, KOPL, pre-trains phoneme and word encoders with a contrastive objective and then fine-tunes downstream models on phoneme, word, and mixed inputs, averaging their prediction scores. On five Korean NLP tasks, KOPL reports an average improvement of 1.9% over the previous state-of-the-art OOV method on test subsets containing OOV words, and it can be added to existing static and contextual Korean embedding models without retraining them.

What carries the argument

The load-bearing mechanism is the joint pre-training of a phoneme encoder $F_P$ and a word encoder $F_W$ so that the interpolated representation $F(w)=\lambda F_P(p_w)+(1-\lambda)F_W(w)$ predicts the pre-trained embedding $v_w$ under a contrastive loss. This makes the two encoders complementary: the phoneme path captures pronunciation shared by words with similar sounds, and the word path preserves spelling-based distinctions. During fine-tuning, three losses over phoneme, word, and mixed inputs train the downstream model, and a weighted average of the three prediction scores is used at inference. The interpolation ratio $\lambda=0.1$ and equal loss weights $\alpha_1=\alpha_2=\alpha_3=1$ and ensemble weights $\beta_1=\beta_2=\beta_3=1/3$ are the specific settings in the reported experiments.

What would settle it

Take a set of Korean OOV words whose IPA conversions from hangul_to_ipa are known to be wrong or ambiguous (for example, foreign loanwords with irregular pronunciation or deliberately misspelled words where pronunciation differs from spelling) and run KOPL with those phoneme inputs. If performance on that set is no better than the word-only baseline, the claim that phoneme representations carry useful signal fails.

Watch

Extended reading notes

Core claim

The central claim is that the correlation between Hangul letters and phonemes is a usable signal for representing Korean OOV words. KOPL trains two small self-attention encoders, one over IPA phoneme sequences and one over mixed subcharacter and morpheme inputs, to predict pre-trained word embeddings through a contrastive loss, with a linear interpolation of the two representations as the prediction target. At fine-tuning time, the downstream model is fed phoneme, word, and mixed representations as three parallel inputs, trained on the sum of three losses, and at inference the prediction scores of the three modalities are averaged. The paper reports that this consistently outperforms comparable English-oriented OOV methods adapted to Korean, as well as the best previous Korean-specific methods, on KOLD, KLUE-TC, NSMC, KLUE-DP, and KLUE-NER, and that removing any of the three input types degrades performance. The ablation results are offered as evidence that phoneme information, word information, and their connection each contribute.

Load-bearing premise

The whole method assumes the Hangul-to-IPA conversion is accurate and informative for every OOV word, including misspellings and loanwords; if the converter produces wrong or unstable phoneme sequences, the phoneme path adds noise and the reported gains can disappear.

Editorial extensions

If this is right

  • KOPL can be layered onto existing static embedding models such as SISG(jm) and SISG(BTS) and onto contextual models such as BERT(Morph.), improving both OOV and original test subsets in every reported dataset.
  • The ablation study implies the phoneme path is not redundant: without phoneme inputs, the model loses roughly 2-3 points on OOV subsets; without word inputs it loses accuracy; without mixed inputs performance drops the most.
  • Because KOPL needs only pre-trained embeddings and a phoneme conversion tool, it can improve Korean NLP systems without re-training the base embedding models.
  • The cross-modal ensemble beats an ensemble of three LOVE checkpoints, suggesting the gain comes from phoneme information, not just ensembling.

Reading between the lines

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

  • The same design may transfer to other phonemic or featural writing systems, but the gains depend on the quality and stability of orthography-to-phoneme conversion, which is not analyzed in the paper for irregular forms.
  • KOPL could be tested as a lightweight robustness layer for misspellings and stylized internet Korean, since its motivating example is a misspelling recovered via pronunciation.
  • A testable extension is to compare KOPL against an oracle IPA conversion and against a deliberately corrupted conversion to quantify how much of the gain rides on conversion accuracy.
  • For languages with less systematic spelling-pronunciation mapping, the interpolation weight $\lambda$ may need to be learned per language rather than fixed at 0.1.
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. This paper introduces KOPL, a framework that learns phoneme and word representations for Korean out-of-vocabulary words. KOPL first pre-trains two modules, FP and FW, on in-vocabulary words: FP consumes IPA phoneme sequences produced by the external hangul_to_ipa tool, FW consumes subcharacter/morpheme sequences, and their interpolation F(w)=λFP(p_w)+(1−λ)FW(w) is trained with a contrastive loss to predict pre-trained SISG(jm) embeddings. During downstream fine-tuning, the authors feed phoneme, word, and mixed representations as parallel inputs into a task model with three losses, and average the prediction scores at inference. Experiments on five Korean tasks (KOLD, KLUE-TC, NSMC, KLUE-DP, KLUE-NER) compare against Korean and English OOV baselines, report OOV-subset and original-set scores, include an ablation of input types, and demonstrate plug-and-play combination with SISG(jm), SISG(BTS), and BERT(Morph.). The main reported result is an average 1.9% improvement over the prior state of the art SISG(BTS) across the five tasks.

Significance. The paper addresses a genuine and language-specific problem, and the idea of using IPA phoneme sequences as auxiliary input for Korean OOV words is novel and cleanly motivated by the phonemic nature of Hangul. The experimental design is broad: five diverse tasks, several baselines (including Korean-specific and English-derived methods), an ensemble control, an ablation study, and adaptability experiments with both static and contextual models. The authors also release code, which supports reproducibility. If the claimed gains are confirmed, KOPL would provide a simple, plug-and-play component that can be added to existing Korean embedding models. However, two load-bearing gaps currently limit the strength of the claim: the phonemizer used to generate the central input is not validated on actual OOV words, and the reported gains are not accompanied by variance or significance testing. The result is therefore promising but not yet fully established.

major comments (4)
  1. [§2.1, footnote 4; §3.1] The central novelty of KOPL is the IPA phoneme pathway, but the paper never validates the outputs of the external hangul_to_ipa program on the words that actually appear in evaluation. The OOV subsets are by construction enriched for rare words, misspellings, and loanwords, and the motivating example in §1 is the misspelling '마싯다'; these are exactly the cases where grapheme-to-phoneme conversion can fail (e.g., tensification and other phonological processes in '맛있다' → [maɕit͈a]). If hangul_to_ipa systematically mis-transcribes such inputs, FP receives noisy input and the improvement seen in Table 3 cannot be attributed to phonemic content. Please provide a validation of the phonemizer on the OOV test vocabulary (for example, error rate against a standard Korean pronunciation dictionary, or a sampled list of IPA outputs with human assessment), and discuss failure cases. This is load-bearing because the ablation in Table 4 only shows that removing the standalone phoneme input hurts; it does not show that the phoneme content, rather than the additional input stream, causes the gain.
  2. [§4.1, Table 3] The claim that KOPL 'significantly improves' or 'consistently attains higher' performance is not backed by any measure of uncertainty. Table 3 reports single runs for each model; no standard deviations, confidence intervals, or significance tests are given, and the OOV subsets are small (0.5k–8.5k examples). Several per-task differences are small (e.g., +0.5 on KLUE-TC OOV, +0.4 on KLUE-TC original), which could easily fall within seed-level noise. Please report means and standard deviations over multiple runs, apply a paired test for the key KOPL vs. SISG(BTS) and KOPL vs. Ensemble comparisons, and adjust the 'significant' language to match the statistical evidence.
  3. [§4.3, Table 5 vs Table 3] The baseline SISG(jm) is reported inconsistently across the two tables: in Table 3 the KLUE-TC OOV score is 77.9, while in Table 5 it is 76.1. Because the model and dataset should be identical, the mismatch makes the improvement of SISG(jm)+KOPL over SISG(jm) in Table 5 non-comparable for that cell. Please either rerun the same baseline for both tables or explain the discrepancy (e.g., different seeds or preprocessing).
  4. [§4.2, Table 4] The ablation is difficult to interpret because the 'mixed' representation defined in Eq. 1 already contains both phoneme and word information: F(w) = λFP(p_w) + (1−λ)FW(w). Therefore row 6 ('word + mixed') does not 'omit phoneme representations' as the text claims; it still includes phoneme information through the mixed stream, and row 5 similarly still contains word information. Only comparisons such as row 2 vs. row 4 cleanly isolate the effect of adding the standalone phoneme input. The conclusion that 'each type of representation contains distinct information' is not supportable from this design. Please either redesign the ablation so that the input types are mutually exclusive, or reinterpret the rows as ablating only the standalone input streams and qualify the discussion accordingly.
minor comments (6)
  1. [Title and abstract] The title and abstract contain 'Out-of-V ocabulary' with an unintended space; please correct this typo.
  2. [Figure 1 and §2.1] The IPA notation such as [mAsitt'A] uses capital letters and an apostrophe that are not standard IPA; please either switch to standard IPA symbols or explain the transcription convention.
  3. [Footnote 5] If 'similar pronunciation' is meant to be quantified via edit distance, please state this explicitly and report the edit-distance threshold used, since none is given.
  4. [§3.3] The sentence 'we straightforwardly use a list of IPA letters' should clarify whether the sequence is the raw output of hangul_to_ipa and how the program handles spaces or multi-syllable words.
  5. [§3.2] The description of the Ensemble baseline is brief; please state whether the same downstream architecture and hyperparameters were used and whether the three LOVE checkpoints are the same ones used for LOVE in Table 3.
  6. [Abstract and §3.1] The abstract says 'real-world OOV examples,' but the OOV subsets are constructed by absence from the SISG(jm) vocabulary; please clarify this in the abstract or methodology.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KOPL's downstream gains are measured on external tasks, not derived from its SISG(jm) training target.

full rationale

KOPL pre-trains the phoneme and word encoders FP and FW to predict SISG(jm) word embeddings via the contrastive loss in Eq. 2. This makes SISG(jm) a teacher signal, but it is not the claimed result. The central claim is improved performance on five external Korean NLP benchmarks (KOLD, KLUE-TC, NSMC, KLUE-DP, KLUE-NER), which are labeled independently and evaluated with standard metrics. Nothing in the training objective forces downstream gains over SISG(jm); in fact, the ablation study (Table 4) and the Ensemble baseline show that the phoneme pathway and mixed representations contribute beyond the teacher, and the ensemble weights in Eq. 6 are fixed rather than fitted to the evaluation sets. The paper does not define any predicted quantity in terms of a fitted parameter that trivially recovers the result. The only notable dependency is the external hangul_to_ipa phonemizer, whose accuracy on OOV words is not validated; this is a correctness and robustness concern, not circularity, because errors in the phoneme input would weaken or add noise to the method rather than force the reported improvement. Self-citations such as [9] are used as baselines and descriptive motivation about Korean subword units, not as load-bearing justification for KOPL's novelty. The derivation chain is therefore self-contained with respect to the claimed contribution.

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

The method relies on SISG(jm) pre-trained embeddings, the hangul_to_ipa tool, and standard contrastive learning. No new entities are introduced, but several domain assumptions are load-bearing.

free parameters (4)
  • mixup ratio lambda = 0.1
    Hand-chosen to balance phoneme and word representations in Eq. 1; no sensitivity analysis reported.
  • temperature tau = 0.07
    Standard contrastive learning temperature, used in prior work [4,16].
  • loss weights alpha1, alpha2, alpha3 = 1, 1, 1
    Described as tuned hyperparameters in Eq. 5, then set to equal values; no search reported.
  • ensemble weights beta1, beta2, beta3 = 1/3 each
    Fixed equal weights for ensembling in Eq. 6.
assumptions (4)
  • domain assumption Hangeul is a phonemic script with high correlation between letters and phonemes.
    Used in Introduction and Section 2 to justify using phoneme representations.
  • domain assumption The hangul_to_ipa program produces correct IPA sequences for Korean words.
    Section 2.1: phoneme input is constructed by IPA symbols from this tool; incorrect conversion would make phoneme representations unreliable.
  • domain assumption Pre-trained word embeddings vw from SISG(jm) are a valid supervision target.
    Section 2.1: contrastive loss predicts vw; if these embeddings are poor, learned representations inherit their defects.
  • standard math Contrastive loss with in-batch negatives is a sufficient objective.
    Borrowed from LOVE/mimicking; not proven here but a standard technique.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Handling Korean Out-of-Vocabulary Words with Phoneme Representation Learning." pith.science (2026). https://pith.science/paper/GOX5KE6H

@misc{pith2026250704018,
  author       = {Pith},
  title        = {Pith review of: Handling Korean Out-of-Vocabulary Words with Phoneme Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GOX5KE6H}},
  note         = {Machine review of arXiv:2507.04018}
}
read the original abstract

In this study, we introduce KOPL, a novel framework for handling Korean OOV words with Phoneme representation Learning. Our work is based on the linguistic property of Korean as a phonemic script, the high correlation between phonemes and letters. KOPL incorporates phoneme and word representations for Korean OOV words, facilitating Korean OOV word representations to capture both text and phoneme information of words. We empirically demonstrate that KOPL significantly improves the performance on Korean Natural Language Processing (NLP) tasks, while being readily integrated into existing static and contextual Korean embedding models in a plug-and-play manner. Notably, we show that KOPL outperforms the state-of-the-art model by an average of 1.9%. Our code is available at https://github.com/jej127/KOPL.git.

Figures

Figures reproduced from arXiv: 2507.04018 by the authors.

Figure 1
Figure 1. Illustration of humans’ forming the word meaning. When encountering the un￾seen word “*마싯다[mAsitt’A]*dilisious”, hu￾mans can recall its phoneme information “[mAsitt’A]” to infer the correct word “맛있 다[mAsitt’A]delicious”. For instance, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of KOPL. The meanings of inputs “ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 36 canonical work pages

  1. [1]

    Cambridge University Press (1999)

    Association, I.P.: Handbook of the International Phonetic Association: A guide to the use of the International Phonetic Alphabet. Cambridge University Press (1999)

    work page 1999

  2. [2]

    Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with subword information. Trans. Assoc. Comput. Linguistics 5, 135–146 (2017)

    work page 2017

  3. [3]

    Thomson Brooks/Cole Publishing Co (1986)

    Carroll, D.W.: Psychology of language. Thomson Brooks/Cole Publishing Co (1986)

    work page 1986

  4. [4]

    In: ACL 2022

    Chen, L., Varoquaux, G., Suchanek, F.M.: Imputing out-of-vocabulary embeddings with LOVE makes languagemodels robust with little cost. In: ACL 2022. pp. 3488–3504 (2022)

    work page 2022

  5. [5]

    In: ACL 2019

    Hu, Z., Chen, T., Chang, K., Sun, Y .: Few-shot representation learning for out-of-vocabulary words. In: ACL 2019. pp. 4102–4112 (2019)

    work page 2019

  6. [6]

    CoRR abs/1508.01991 (2015)

    Huang, Z., Xu, W., Yu, K.: Bidirectional LSTM-CRF models for sequence tagging. CoRR abs/1508.01991 (2015)

    arXiv 2015

  7. [7]

    In: EMNLP 2022

    Jeong, Y ., Oh, J., Lee, J., Ahn, J., Moon, J., Park, S., Oh, A.: KOLD: korean offensive lan- guage dataset. In: EMNLP 2022. pp. 10818–10833 (2022)

    work page 2022

  8. [8]

    In: AAAI 2020

    Jin, D., Jin, Z., Zhou, J.T., Szolovits, P.: Is BERT really robust? A strong baseline for natural language attack on text classification and entailment. In: AAAI 2020. pp. 8018–8025 (2020)

    work page 2020

Show all 38 references
  1. [9]

    In: EMNLP 2022

    Kim, N., Park, J., Choi, J., Jeon, E., Kang, Y ., Lee, S.: Break it down into BTS: basic, tiniest subword units for korean. In: EMNLP 2022. pp. 7007–7024 (2022)

    2022

  2. [10]

    In: Findings of the Association for Computational Linguistics ACL 2024

    Kim, S., Park, J., Kim, Y ., Lee, S.: KOMBO: Korean character representations based on the combination rules of subcharacters. In: Findings of the Association for Computational Linguistics ACL 2024. pp. 5102–5119 (2024)

    2024

  3. [11]

    In: EACL 2021

    Kwon, O., Kim, D., Lee, S., Choi, J., Lee, S.: Handling out-of-vocabulary problem in hangeul word embeddings. In: EACL 2021. pp. 3213–3221 (2021)

    2021

  4. [12]

    In: Find- ings of the Association for Computational Linguistics ACL 2024

    Lee, J., Moon, H., Lee, S., Park, C., Eo, S., Ko, H., Seo, J., Lee, S., Lim, H.: Length-aware byte pair encoding for mitigating over-segmentation in Korean machine translation. In: Find- ings of the Association for Computational Linguistics ACL 2024. pp. 2287–2303 (2024)

    2024

  5. [13]

    In: EMNLP

    Li, J., Wang, Q., Mao, Z., Guo, J., Yang, Y ., Zhang, Y .: Improving chinese spelling check by character pronunciation prediction: The effects of adaptivity and granularity. In: EMNLP

  6. [14]

    In: IJCAI 2018

    Liang, B., Li, H., Su, M., Bian, P., Li, X., Shi, W.: Deep text classification can be fooled. In: IJCAI 2018. pp. 4208–4215 (2018)

    2018

  7. [15]

    In: ACL 2023

    Liang, Z., Quan, X., Wang, Q.: Disentangled phonetic representation for chinese spelling correction. In: ACL 2023. pp. 13509–13521 (2023)

    2023

  8. [16]

    In: ACL 2023

    Liang, Z., Lu, Y ., Chen, H., Rao, Y .: Graph-based relation mining for context-free out-of- vocabulary word embedding learning. In: ACL 2023. pp. 14133–14149 (2023)

    2023

  9. [17]

    In: AACL/IJCNLP 2020

    Park, K., Lee, J., Jang, S., Jung, D.: An empirical study of tokenization strategies for various korean NLP tasks. In: AACL/IJCNLP 2020. pp. 133–142 (2020) 12 Kim et al

    2020

  10. [18]

    In: ACL 2018

    Park, S., Byun, J., Baek, S., Cho, Y ., Oh, A.: Subword-level word vector representations for korean. In: ACL 2018. pp. 2429–2438 (2018)

    2018

  11. [19]

    In: NeurIPS 2021 (2021)

    Park, S., Moon, J., Kim, S., Cho, W., Han, J., Park, J., Song, C., Kim, J., Song, Y ., Oh, T.H., Lee, J., Oh, J., Lyu, S., Jeong, Y ., Lee, I., Seo, S., Lee, D., Kim, H., Lee, M., Jang, S., Do, S., Kim, S., Lim, K., Lee, J., Park, K., Shin, J., Kim, S., Park, E.L., Oh, A., Ha,...

    2021

  12. [20]

    In: Proceedings of the Second Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature

    Pilán, I., V olodina, E.: Exploring word embeddings and phonological similarity for the unsu- pervised correction of language learner errors. In: Proceedings of the Second Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and...

    2018

  13. [21]

    In: EMNLP 2017

    Pinter, Y ., Guthrie, R., Eisenstein, J.: Mimicking word embeddings using subword rnns. In: EMNLP 2017. pp. 102–112 (2017)

    2017

  14. [22]

    London, UK: H utchinson (1985)

    Sampson, G.: Writing systems. London, UK: H utchinson (1985)

    1985

  15. [23]

    In: NAACL-HLT 2019

    Sasaki, S., Suzuki, J., Inui, K.: Subword-based compact reconstruction of word embeddings. In: NAACL-HLT 2019. pp. 3498–3508 (2019)

    2019

  16. [24]

    In: NAACL-HLT 2019

    Schick, T., Schütze, H.: Attentive mimicking: Better word embeddings by attending to infor- mative contexts. In: NAACL-HLT 2019. pp. 489–494 (2019)

    2019

  17. [25]

    In: AAAI 2019

    Schick, T., Schütze, H.: Learning semantic representations for novel words: Leveraging both form and context. In: AAAI 2019. pp. 6965–6973 (2019)

    2019

  18. [26]

    In: EMNLP 2023

    Seo, J., Moon, H., Lee, J., Eo, S., Park, C., Lim, H.: CHEF in the language kitchen: A generative data augmentation leveraging Korean morpheme ingredients. In: EMNLP 2023. pp. 6014–6029 (2023)

    2023

  19. [27]

    In: Proceedings of the Fifteenth Workshop on Computational Research in Phonetics, Phonology, and Morphology

    Sofroniev, P., Çöltekin, Ç.: Phonetic vector representations for sound sequence alignment. In: Proceedings of the Fifteenth Workshop on Computational Research in Phonetics, Phonology, and Morphology. Brussels, Belgium (2018)

    2018

  20. [28]

    In: ACL 2022

    Su, H., Shi, W., Shen, X., Zhou, X., Ji, T., Fang, J., Zhou, J.: Rocbert: Robust chinese bert with multimodal contrastive pretraining. In: ACL 2022. pp. 921–931 (2022)

    2022

  21. [29]

    arXiv preprint arXiv:2003.04985 (2020)

    Sun, L., Hashimoto, K., Yin, W., Asai, A., Li, J., Yu, P.S., Xiong, C.: Adv-bert: BERT is not robust on misspellings! generating nature adversarial samples on BERT. arXiv preprint arXiv:2003.04985 (2020)

    2020 arXiv

  22. [30]

    In: Interspeech 2021

    Sundararaman, M.N., Kumar, A., Vepa, J.: Phonemebert: Joint language modelling of phoneme sequence and ASR transcript. In: Interspeech 2021. pp. 3236–3240 (2021)

    2021

  23. [31]

    Üstün, A., Kurfali, M., Can, B.: Characters or morphemes: How to represent words? In: Proceedings of The Third Workshop on Representation Learning for NLP, Rep4NLP@ACL

  24. [32]

    In: NIPS 2017

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NIPS 2017. pp. 5998–6008 (2017)

    2017

  25. [33]

    In: EMNLP 2016

    Wieting, J., Bansal, M., Gimpel, K., Livescu, K.: Charagram: Embedding words and sen- tences via character n-grams. In: EMNLP 2016. pp. 1504–1515 (2016)

    2016

  26. [34]

    In: Findings of the Association for Com- putational Linguistics: ACL/IJCNLP 2021

    Zhang, R., Pang, C., Zhang, C., Wang, S., He, Z., Sun, Y ., Wu, H., Wang, H.: Correcting chinese spelling errors with phonetic pre-training. In: Findings of the Association for Com- putational Linguistics: ACL/IJCNLP 2021. pp. 2250–2261 (2021)

    2021

  27. [35]

    In: EMNLP 2018

    Zhao, J., Mudgal, S., Liang, Y .: Generalizing word embeddings using bag of subwords. In: EMNLP 2018. pp. 601–606 (2018)

    2018

  28. [36]

    In: NAACL 2024

    Zhu, J., Yang, C., Samir, F., Islam, J.: The taste of IPA: towards open-vocabulary keyword spotting and forced alignment in any language. In: NAACL 2024. pp. 750–772 (2024)

    2024

  29. [37]

    In: NAACL-HLT 2019

    Zhu, Y ., Vulic, I., Korhonen, A.: A systematic study of leveraging subword information for learning word representations. In: NAACL-HLT 2019. pp. 912–932 (2019)

    2019

  30. [2022]

    4275–4286 (2022)

    pp. 4275–4286 (2022)

    2022

Pith tools

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