Pith. sign in

REVIEW 4 major objections 4 minor 51 references

BnBERT-iPET: Sparse Few-Shot Language Modeling for Bengali via Lottery Ticket Pruning

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

Pith's one-line read A Bengali language model reduced to 10% of its weights can still match far larger models on several NLP tasks.

desk verdict A worthwhile combination of iPET and lottery-ticket pruning for Bengali, but impossible F1 scores in the key tables undermine the central claim. read the letter →

arxiv 2608.05104 v1 pith:UQOKNHA7 submitted 2026-08-05 cs.LG

classification cs.LG
keywords BengaliNLPfew-shotlearningiterativepatternexploitingtraininglotterytickethypothesisnetworkpruninglow-resourcelanguagemodelingtextclassificationtoken
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 show that a very small Bengali language model, with 90% of its connections removed, can perform about as well as much larger models on downstream tasks. The authors build a tiny BERT-style model, make it a few-shot learner with iterative pattern exploiting training (iPET), then prune it with the Lottery Ticket Hypothesis until only 10% of the original edges remain. They evaluate this pruned model on sentiment, emotion, news categorization, authorship attribution, POS tagging, and punctuation restoration, and report that it matches or beats large multilingual baselines such as Bangla Electra, Indic-BERT, and XLM-RoBERTa on several benchmarks while coming close to BanglaBERT on some. If correct, the result would make competitive Bengali NLP practical on machines with little memory or compute.

What carries the argument

The argument runs on two mechanisms. iPET converts each task into cloze-style fill-in-the-blank questions through pattern-verbalizer pairs, then iteratively trains generations of models on growing, self-labeled datasets, giving a tiny model few-shot competence. Lottery-ticket pruning then removes 10% of the remaining weights after each round and rewinds the survivors, ending at 90% sparsity; the authors apply it to the fine-tuned model. The base model is small by design, with 4 hidden layers, hidden size 312, and a vocabulary of 30,522 tokens, so the pruned result is light enough for low-resource deployment.

What would settle it

Recompute F1 from the stated precision and recall for every row where all three are reported; Tables VIII and IX already contain mismatches, such as LDC+IITKGP+CRBLP reporting F1=82.01 with P=75.51 and R=74.78 (harmonic mean about 75.1), and YouTube comments reporting P=70.29, R=57.43, F1=58.76 (harmonic mean about 63.2). A systematic mismatch would show the central comparison is not supported.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that a 90%-sparse few-shot-learned language model can go neck and neck with state-of-the-art models on challenging Bengali tasks. Starting from a tiny ALBERT-style architecture and a custom 80 MB diverse Bengali corpus, the authors produce BnBERT, then BnBERT iPET, then BnBERT iPET Pruned, and report that the pruned version, about 10% of the base model's edges, achieves F1 scores comparable to or better than Bangla Electra, Indic-BERT, and XLM-RoBERTa on the six tasks, and approaches BanglaBERT on some of them.

Load-bearing premise

The claim rests on the reported accuracy numbers being computed and compared fairly; if the precision, recall, and F1 scores in the tables do not actually match each other, the comparison collapses.

Editorial extensions

If this is right

  • A Bengali model with 90% sparsity could be deployed on low-memory devices while keeping accuracy close to much larger models.
  • The pruned model's downstream performance stays competitive even though its language-model perplexity jumps to 46.86 from 1.05, suggesting task fine-tuning can survive heavy pruning.
  • BanglaBERT still beats the pruned model on most tasks, so the reported claim is one of comparability rather than outright superiority.
  • The inference-time comparison suggests the pruned model is faster than XLM-RoBERTa, supporting the paper's efficiency motivation.
  • The recipe of a tiny base model, iPET for few-shot learning, and lottery-ticket pruning to 90% sparsity is presented as a way to reduce the computational cost of Bengali NLP.

Reading between the lines

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

  • The paper does not vary the sparsity level, so it leaves open whether 90% is the best operating point or whether a different keep ratio gives a better accuracy-compute tradeoff.
  • If the approach transfers, similar pruned few-shot models could be built for other low-resource languages by fine-tuning a tiny multilingual checkpoint, applying iPET, and pruning, an extension the authors do not test.
  • The paper describes pruning during MLM training and then evaluates on downstream tasks, so it does not isolate whether finding the lottery ticket during language-model training, during downstream fine-tuning, or at some intermediate stage changes the result.
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 BnBERT, BnBERT-iPET, and BnBERT-iPET Pruned, three Bengali language models built by fine-tuning a tiny Chinese ALBERT checkpoint on an 80 MB mixed Bengali corpus. The authors apply iterative pattern exploiting training (iPET) for few-shot learning and lottery-ticket pruning to reach 90% sparsity. They report perplexity and evaluations on six downstream tasks (emotion, authorship, news, POS tagging, sentiment, punctuation restoration) and compare against BanglaBERT, Bangla Electra, Indic-BERT, XLM-RoBERTa, and other baselines. The central claim is that the 90%-sparse BnBERT-iPET Pruned model, retaining roughly 10% of the weights, is competitive with much larger state-of-the-art models on these Bengali benchmarks.

Significance. If the empirical claims were reliable, the paper would be a practically valuable demonstration that a tiny, highly sparse, few-shot-trained transformer can approach the performance of large multilingual models in a low-resource language. The authors have compiled a new corpus, applied iPET to Bengali, and evaluated a broad set of tasks, which is a substantial amount of work. However, the paper is purely empirical and ships no code, data, or evaluation scripts. More importantly, the reported evaluation tables contain multiple internally inconsistent F1 values, and one of these impossible values is the exact number used in the POS-tagging comparison that supports the headline claim. The central result is therefore not currently supported by the paper's own data.

major comments (4)
  1. [Tables V, VII, VIII; Section V-D] Section V states that precision, recall, and F1 are weighted averages. Under support-weighted averaging, per-class F1 is at most the arithmetic mean of per-class precision and recall, so the reported F1 can never exceed (P+R)/2. Several rows violate this elementary bound. In Table VIII, the LDC+IITKGP+CRBLP row for BnBERT iPET Pruned reports P=75.51, R=74.78, F1=82.01, while (P+R)/2=75.15. In Table VII, BnBERT iPET on the News set reports P=87.94, R=95.20, F1=93.76, with an upper bound of 91.57; the Reference and ASR rows violate the same bound. In Table V, BanglaBERT reports P=97.71, R=97.88, F1=98.71, while (P+R)/2=97.80. In Table VIII, Bangla Electra on the merged POS set reports P=75.1, R=73.0, F1=74.7, with an upper bound of 74.05. Section V-D explicitly uses the impossible F1=82.01 to argue that BnBERT iPET Pruned is superior to Bangla Electra and DistilBERT-m on POS tagging. Because this is a load-bearing comparison and the violations are widespread, the evaluation tables do not support the paper's central claim.
  2. [Section III-B; Table II] The paper defines a matching subnetwork and a winning ticket in Eq. (3), but no experiment verifies this condition. Table II reports training-loss perplexity of 46.86 for the pruned model versus 3.87 for BnBERT and 1.05 for BnBERT iPET, so under the MLM objective the pruned model is far from matching the dense model. Downstream accuracy on some tasks does not establish that the pruned subnetwork is a winning ticket in the sense defined by the paper. The claim that the 90%-sparse model is a lottery ticket therefore remains unsubstantiated.
  3. [Sections III-C and IV-C] The base model is initialized from the tiny Chinese ALBERT checkpoint, but the paper never describes how the vocabulary and tokenizer are adapted from Chinese to Bengali. A model with a Chinese wordpiece vocabulary cannot process Bengali text unless the vocabulary is replaced or augmented, and any replacement changes the embedding layer in ways that are not reported. This omission is essential for the validity of the pretrained initialization and for reproducibility.
  4. [Section V; Tables IV-IX] All comparison numbers for Bangla Electra, Indic-BERT, BERT-bn, XLM-RoBERTa, DistilBERT-m, and BanglaBERT are quoted from prior papers ([12], [27]) rather than measured under the same train/dev/test splits, preprocessing, fine-tuning protocol, and evaluation script used for the authors' models. Given that several quoted baseline rows also violate the F1 bound described above, these numbers cannot be treated as controlled baselines. The authors should rerun all models in a single pipeline and release the evaluation code, data splits, and preprocessing steps.
minor comments (4)
  1. [Table II] Table II is described as "based on training loss," so the perplexity values measure training fit rather than generalization; held-out perplexity should be reported instead.
  2. [Section VI] The inference-time comparison does not specify hardware, framework, batch size, sequence length, or implementation details, so the reported training and inference times are not reproducible.
  3. [Section IV-B] The downstream task datasets are said to come from reference [36], but [36] is the lottery-ticket-for-BERT paper, not a dataset collection; the citation appears to be incorrect.
  4. [General] The paper does not state data/code availability; none of the models, the corpus, or the evaluation scripts are released, which is particularly problematic given the metric inconsistencies in the tables.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical application of published methods, and its central claim rests on measured evaluations rather than on a derivation that returns its own inputs.

full rationale

The paper contains no derivation chain in which a predicted quantity is defined in terms of the target quantity. Equations (1) and (2) are the standard PET conditional-probability definitions taken from Schick and Schütze [22]; equations (3) and (4) are the standard lottery-ticket matching-subnetwork definition from Frankle and Carbin [21] and the standard perplexity formula. The pruning procedure, iPET training, downstream fine-tuning, and perplexity measurements are applied as empirical operations, and the conclusions are based on the resulting tables rather than on any fitted parameter being renamed as a prediction. The only notable concern is that some reported metrics in Tables IV, VIII, and IX appear internally inconsistent, e.g., in Table VIII the LDC+IITKGP+CRBLP row for BnBERT iPET Pruned reports F1=82.01 with precision 75.51 and recall 74.78, whose harmonic mean is approximately 75.1. That is a correctness or evaluation-error issue, not a circularity issue, because the number is not derived from the claim it is used to support. Baseline numbers quoted from references [12] and [27] also may not be matched under an identical protocol, but that is a benchmark-comparability limitation rather than a self-referential derivation. No load-bearing step reduces to its own input, and the paper does not rely on self-citations to justify its central premise. Therefore, no circularity is found.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

This is an empirical machine-learning paper; the central claims rest on the availability of the base pretrained model, the validity of the iPET and LTH procedures as applied, the quality of the unreleased BanglaDDS corpus, and the faithfulness of the comparison protocol. No new theoretical quantities are introduced.

free parameters (3)
  • Sparsity target = 90%
    Chosen by hand as the target sparsity; the central claim is that 10% of edges remain. No ablation over sparsity levels is reported.
  • Pruning fraction per iteration = 10%
    The paper states 'after every iteration, 10 percent of the network is dropped' until 90% sparsity; this schedule is chosen without justification or ablation.
  • iPET generation count and dataset growth factor
    The iPET procedure from [22] requires the number of generations and the constant factor by which the training set grows; the paper does not report these values, though the method depends on them.
assumptions (4)
  • domain assumption Lottery Ticket Hypothesis: iterative magnitude pruning can find trainable subnetworks at 90% sparsity that match the original model's performance.
    Section III-B defines winning tickets but does not verify that the found subnetwork is a 'winning ticket' under Eq. (3); no rewind or mask comparison is reported.
  • domain assumption iPET pattern-verbalizer pairs can be constructed for Bengali tasks and improve few-shot learning.
    Section III-A describes iPET generally, but no patterns or verbalizers are given for any of the six tasks; the authors only state they used 'custom patterns'.
  • ad hoc to paper The tiny Chinese ALBERT model can serve as a base for Bengali.
    Section III-C says the base model starts from ckiplab/albert-tiny-chinese, but Table III reports vocab size 30522 (not the Chinese ALBERT vocab size 21128), and no adaptation of the tokenizer is described.
  • domain assumption Baseline numbers from [12] and [27] are comparable without re-running under identical splits.
    Section V compares directly with quoted baselines; the paper does not establish matching evaluation protocols.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BnBERT-iPET: Sparse Few-Shot Language Modeling for Bengali via Lottery Ticket Pruning." pith.science (2026). https://pith.science/paper/UQOKNHA7

@misc{pith2026260805104,
  author       = {Pith},
  title        = {Pith review of: BnBERT-iPET: Sparse Few-Shot Language Modeling for Bengali via Lottery Ticket Pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UQOKNHA7}},
  note         = {Machine review of arXiv:2608.05104}
}
read the original abstract

Deep neural networks have shown impressive success in NLP tasks owing to their complex structure and huge number of edges. Achieving state-of-the-art performance in natural language processing with a large pre-trained model such as BERT is expensive and time-consuming, carries a large carbon footprint, and is difficult to realize on machines with minimal computational capability. This creates a barrier to training complex models for resource-constrained languages such as Bengali. However, in a complex neural model, not all edges are equally impactful, and the contributions of some of them can be neglected. Pruning promises to reduce the memory footprint of regular networks, shorten the training time of ever-growing networks, and increase inference efficiency without sacrificing comparable performance. In this work, we introduce BnBERT-iPET, a sparse few-shot language modeling approach for Bengali, and experimentally show that a lightweight few-shot-learned language model retaining only 10% of the edges of an initial model such as BERT can perform neck and neck with much larger models on challenging tasks for a resource-constrained language such as Bengali. By learning from few shots through iterative pattern exploiting training and achieving 90% sparsity with the Lottery Ticket Hypothesis pruning technique, our pruned BnBERT-iPET model proves to be a tough competitor to state-of-the-art language models such as Bangla Electra, Indic-BERT, and XLM-RoBERTa on downstream tasks over standard benchmark datasets of the Bengali language.

Figures

Figures reproduced from arXiv: 2608.05104 by the authors.

Figure 1
Figure 1. Illustration of iterative pattern exploiting training and pruning of a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An illustration of the work process. The performance of each model [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The detailed process of data collection and construction of a diverse [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Train, test, and development data distributions for the emotion dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Train, test, and development data distributions for the authorship [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: Train, test, and development data distributions for the YouTube [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 9
Figure 9. Figure 9: Train, test, and development data distributions for the SAIL dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 12
Figure 12. Figure 12: Train, test, and development data distributions for the CogniSenti [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 13
Figure 13. Figure 13: Train, test, and development data distributions for the combined [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 38 canonical work pages

  1. [12]

    A Review of Bangla Natural Language Processing Tasks and the Utility of Transformer Models

    Alam, F., Hasan, A., Alam, T., Khan, A., Tajrin, J., Khan, N. and Chowd- hury, S.A., 2021. A review of bangla natural language processing tasks and the utility of transformer models. arXiv preprint arXiv:2107.03844

  2. [27]

    and Shahriyar, R., 2022

    Bhattacharjee, A., Hasan, T., Uddin, W.A., Mubasshir, K., Islam, M.S., Iqbal, A., Rahman, M.S. and Shahriyar, R., 2022. BanglaBERT: Lan- guage model pretraining and benchmarks for low-resource language understanding evaluation in bangla. Findings of the North American Chapter of the Association for Computational Linguistics: NAACL

  3. [1]

    and Manning, C.D., 2015

    Hirschberg, J. and Manning, C.D., 2015. Advances in natural language processing. Science, 349(6245), pp.261-266. PREPRINT 13

  4. [2]

    and Kaur, R., 2021

    Ranathunga, S., Lee, E.S.A., Skenduli, M.P., Shekhar, R., Alam, M. and Kaur, R., 2021. Neural machine translation for low-resource languages: A survey. arXiv preprint arXiv:2106.15115

  5. [3]

    Language models: past, present, and future

    Li, H., 2022. Language models: past, present, and future. Communica- tions of the ACM, 65(7), pp.56-63

  6. [4]

    An unsupervised parts-of-speech tagger for the bangla language

    Ali, H., 2010. An unsupervised parts-of-speech tagger for the bangla language. Department of Computer Science, University of British Columbia, 20, pp.1-8

  7. [5]

    Research on Bangla language processing in Bangladesh: progress and challenges

    Islam, M.S., 2009, June. Research on Bangla language processing in Bangladesh: progress and challenges. In 8th international language & development conference (pp. 23-25)

  8. [6]

    and Khan, M., 2007

    Hasan, F.M., UzZaman, N. and Khan, M., 2007. Comparison of differ- ent POS Tagging Techniques (N-Gram, HMM and Brill’s tagger) for Bangla. In Advances and innovations in systems, computing sciences and software engineering (pp. 121-126). Springer, Dordrecht

Show all 51 references
  1. [7]

    and Bandyopadhyay, S., 2010

    Ekbal, A. and Bandyopadhyay, S., 2010. Named entity recognition using support vector machine: A language independent approach. International Journal of Electrical and Computer Engineering, 4(3), pp.589-604

  2. [8]

    and Izhar, M.N., 2013

    Hasan, K.A., Islam, S., Mashrur-E-Elahi, G.M. and Izhar, M.N., 2013. Sentiment recognition from bangla text. In Technical Challenges and Design Issues in Bangla Language Processing (pp. 315-327). IGI Global

  3. [9]

    and Bandyopadhyay, S., 2010

    Das, A. and Bandyopadhyay, S., 2010. Sentiwordnet for bangla. Knowl- edge Sharing Event-4: Task, 2, pp.1-8

  4. [10]

    and Sarker, I.H., 2021

    Das, A., Sharif, O., Hoque, M.M. and Sarker, I.H., 2021. Emotion classification in a resource constrained language using transformer-based approach. arXiv preprint arXiv:2104.08613

  5. [11]

    and Hasan, T., 2021

    Sen, O., Fuad, M., Islam, M.D., Rabbi, J., Hasan, M.D., Baz, M., Masud, M., Awal, M., Fime, A.A., Fuad, M. and Hasan, T., 2021. Bangla Natural Language Processing: A Comprehensive Review of Classical, Machine Learning, and Deep Learning Based Methods. arXiv preprint arXiv:2105.14875

  6. [14]

    and Stoyanov, V ., 2019

    Liu, Y ., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L. and Stoyanov, V ., 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

  7. [15]

    and Wolf, T., 2019

    Sanh, V ., Debut, L., Chaumond, J. and Wolf, T., 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108

  8. [16]

    and Chen, Y ., 2015, June

    Chen, W., Wilson, J., Tyree, S., Weinberger, K. and Chen, Y ., 2015, June. Compressing neural networks with the hashing trick. In International conference on machine learning (pp. 2285-2294). PMLR

  9. [17]

    Gradient and Magnitude Based Pruning for Sparse Deep Neural Networks

    Belay, K., 2022. Gradient and Magnitude Based Pruning for Sparse Deep Neural Networks

  10. [18]

    and Graf, H.P., 2016

    Li, H., Kadav, A., Durdanovic, I., Samet, H. and Graf, H.P., 2016. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710

  11. [19]

    and Sun, J., 2017

    He, Y ., Zhang, X. and Sun, J., 2017. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE international conference on computer vision (pp. 1389-1397)

  12. [20]

    and Dally, W.J., 2015

    Han, S., Mao, H. and Dally, W.J., 2015. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149

  13. [21]

    and Carbin, M., 2018

    Frankle, J. and Carbin, M., 2018. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635

  14. [22]

    and Sch ¨utze, H., 2020

    Schick, T. and Sch ¨utze, H., 2020. It’s not just size that matters: Small language models are also few-shot learners. arXiv preprint arXiv:2009.07118

  15. [23]

    and Jana, A., Sentiment Analysis For Bengali Using Transformer Based Models

    Bhowmick, A. and Jana, A., Sentiment Analysis For Bengali Using Transformer Based Models

  16. [24]

    and Alam, F., 2020, November

    Alam, T., Khan, A. and Alam, F., 2020, November. Punctuation restora- tion using transformer models for high-and low-resource languages. In Proceedings of the Sixth Workshop on Noisy User-generated Text (W- NUT 2020) (pp. 132-142)

  17. [25]

    and Sch ¨utze, H., 2020

    Schick, T. and Sch ¨utze, H., 2020. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676

  18. [26]

    and Koshiba, T., 2022

    Kowsher, M., Sami, A.A., Prottasha, N.J., Arefin, M.S., Dhar, P.K. and Koshiba, T., 2022. Bangla-BERT: Transformer-based Efficient Model for Transfer Learning and Language Understanding. IEEE Access

  19. [28]

    and Gupta, S., 2017

    Zhu, M. and Gupta, S., 2017. To prune, or not to prune: explor- ing the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878

  20. [29]

    and Lei, T., 2019

    Wang, Z., Wohlwend, J. and Lei, T., 2019. Structured pruning of large language models. arXiv preprint arXiv:1910.04732

  21. [30]

    and Toutanova, K., 2018

    Devlin, J., Chang, M.W., Lee, K. and Toutanova, K., 2018. Bert: Pre- training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

  22. [31]

    and Keutzer, K., 2020, April

    Shen, S., Dong, Z., Ye, J., Ma, L., Yao, Z., Gholami, A., Mahoney, M.W. and Keutzer, K., 2020, April. Q-bert: Hessian based ultra low precision quantization of bert. In Proceedings of the AAAI Conference on Artificial Intelligence (V ol. 34, No. 05, pp. 8815-8821)

  23. [32]

    and Liu, J., 2019

    Sun, S., Cheng, Y ., Gan, Z. and Liu, J., 2019. Patient knowledge distil- lation for bert model compression. arXiv preprint arXiv:1908.09355

  24. [33]

    and Morcos, A.S., 2019

    Yu, H., Edunov, S., Tian, Y . and Morcos, A.S., 2019. Playing the lottery with rewards and multiple languages: lottery tickets in rl and nlp. arXiv preprint arXiv:1906.02768

  25. [34]

    and Yosinski, J., 2019

    Zhou, H., Lan, J., Liu, R. and Yosinski, J., 2019. Deconstructing lottery tickets: Zeros, signs, and the supermask. Advances in neural information processing systems, 32

  26. [35]

    and Ganguli, S., 2020

    Tanaka, H., Kunin, D., Yamins, D.L. and Ganguli, S., 2020. Pruning neural networks without any data by iteratively conserving synaptic flow. Advances in Neural Information Processing Systems, 33, pp.6377-6389

  27. [36]

    and Carbin, M., 2020

    Chen, T., Frankle, J., Chang, S., Liu, S., Zhang, Y ., Wang, Z. and Carbin, M., 2020. The lottery ticket hypothesis for pre-trained bert networks. Advances in neural information processing systems, 33, pp.15834- 15846

  28. [37]

    Nafis Irtiza Tripto and Mohammed Eunus Ali. 2018. Detecting multil- abel sentiment and emotions from Bangla youtube comments. In 2018 International Conference on Bangla Speech and Language Processing (ICBSLP). IEEE, 1–6

  29. [38]

    Tanvirul Alam, Akib Khan, and Firoj Alam. 2020. Bangla Text Classi- fication using Transformers. arXiv preprint arXiv:2011.04446 (2020)

  30. [39]

    Aisha Khatun, Anisur Rahman, Md Saiful Islam, et al. 2019. Authorship Attribution in Bangla literature using Character-level CNN. In 2019 22nd International Conference on Computer and Information Technology (ICCIT). IEEE, 1–5

  31. [40]

    Khapra, and Pratyush Kumar

    Anoop Kunchukuttan, Divyanshu Kakwani, Satish Golla, Gokul N.C., Avik Bhattacharyya, Mitesh M. Khapra, and Pratyush Kumar. 2020. AI4Bharat-IndicNLP Corpus: Monolingual Corpora and Word Embed- dings for Indic Languages. arXiv preprint arXiv:2005.00085 (2020)

  32. [41]

    Arid Hasan, Jannatul Tajrin, Shammur Absar Chowdhury, and Firoj Alam

    Md. Arid Hasan, Jannatul Tajrin, Shammur Absar Chowdhury, and Firoj Alam. 2020. Sentiment Classification in Bangla Textual Content: A Comparative Study. In 2020 23rd International Conference on Computer and Information Technology (ICCIT). 1–6

  33. [42]

    Md Rezaul Karim, Bharathi Raja Chakravarthi, Mihael Arcan, John P McCrae, and Michael Cochez. 2020. Classification Bench- marks for Under-resourced Bengali Language based on Multichannel Convolutional-LSTM Network. arXiv (2020), arXiv–2004

  34. [43]

    Braja Gopal Patra, Dipankar Das, Amitava Das, and Rajendra Prasath

  35. [44]

    Md Rahman, Emon Kumar Dey, et al. 2018. Datasets for Aspect-Based Sentiment Analysis in Bangla and Its Baseline Evaluation. Data 3, 2 (2018), 15

  36. [45]

    Sankaran Baskaran, Kalika Bali, Tanmoy Bhattacharya, Pushpak Bhat- tacharyya, Girish Nath Jha, et al. 2008. A common parts-of-speech tagset framework for indian languages. In Proc. of LREC 2008. Citeseer

  37. [46]

    Monojit Choudhury Kalika Bali and Priyanka Biswas. 2010. Indian Lan- guage Part-of-Speech Tagset: Bengali LDC2010T16. Technical Report. Philadelphia: Linguistic Data Consortium

  38. [47]

    A Kumaran. 2007. A Part of Speech Tagger for Indian Languages (POS tagger). Technical Report. Microsoft Research

  39. [48]

    Rabia Sultana Ummi and Fahmina Huda. 2008. Developing language resources for English machine translation. Technical Report. BRAC University

  40. [49]

    Tanvirul Alam, Akib Khan, and Firoj Alam. 2020. Punctuation Restora- tion using Transformer Models for High-and-Low-Resource Languages. In Proceedings of the Sixth Workshop on Noisy User-generated Text (W-NUT 2020). Association for Computational Linguistics, 132–142

  41. [50]

    Saiful Islam, and Ayesha Tasnim

    Aisha Khatun, Anisur Rahman, Hemayet Ahmed Chowdhury, Md. Saiful Islam, and Ayesha Tasnim. 2019. A Subword Level Language Model PREPRINT 14 for Bangla Language. CoRR abs/1911.07613 (2019). arXiv:1911.07613 http://arxiv.org/abs/1911.07613

  42. [51]

    Hugging Face. 2022. ckiplab/albert-tiny-chinese. [Online]. Available: https://huggingface.co/ckiplab/albert-tiny-chinese. Accessed: Aug. 23, 2022

  43. [2015]

    In Mining Intelligence and Knowledge Explo- ration, Rajendra Prasath, Anil Kumar Vuppala, and T

    Shared Task on Sentiment Analysis in Indian Languages (SAIL) Tweets - An Overview. In Mining Intelligence and Knowledge Explo- ration, Rajendra Prasath, Anil Kumar Vuppala, and T. Kathirvalavaku- mar (Eds.). Springer International Publishing, 650–655

Pith tools

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