Cross-lingual Few-shot Learning for Persian Sentiment Analysis with Incremental Adaptation

Farideh Majidi; Ziaeddin Beheshtifard

REVIEW 4 major objections 5 minor 21 references

Cross-lingual Few-shot Learning for Persian Sentiment Analysis with Incremental Adaptation

T0 review · 4 major / 5 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 Fine-tuning multilingual models on a handful of Persian examples reaches 96% accuracy and F1-score for sentiment analysis.

desk verdict Useful empirical matrix for Persian sentiment analysis, but the paper's own tables refute its central incremental-learning claim. read the letter →

arxiv 2507.11634 v1 pith:MCG7PKMZ submitted 2025-07-15 cs.CL cs.AI

Farideh Majidi , Ziaeddin Beheshtifard This is my paper

classification cs.CLcs.AI

keywords sentimentanalysisfew-shotlearningincrementalcross-lingualtransferPersianmultilingualtransformersregularizationlow-resourceNLP

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

Persian sentiment analysis normally suffers from a shortage of labeled data. This paper argues that the gap can be closed by starting from multilingual pre-trained models and adapting them to Persian with very few labeled examples per class. Across settings with 1 to 20 labeled examples per class, drawn from five Persian datasets covering social media, product reviews, food delivery, and book reviews, it reports that mDeBERTa and XLM-RoBERTa reach 96% accuracy and F1-score, while a lighter DistilBERT model falls behind. It also argues that incremental learning, in which datasets are introduced one domain at a time, works best when paired with regularization such as knowledge distillation or rehearsal. If the results hold, low-resource languages could get usable sentiment classifiers without large annotated corpora.

What carries the argument

The load-bearing mechanism is the combination of a multilingual transformer, few-shot sampling, and domain-incremental training with a regularizer. In practice, each of the five Persian datasets is presented to the model sequentially, and one of three mechanisms protects earlier knowledge: Elastic Weight Consolidation penalizes changes to parameters important for previous domains, rehearsal mixes a stored subset of earlier examples into new training, and knowledge distillation lets an earlier version of the model act as a teacher whose soft labels guide current learning. A no-incremental baseline trains on all domains at once, and early stopping with patience three on validation loss guards against overfitting in low-shot conditions. These pieces together are what produce the reported accuracy and F1 numbers.

What would settle it

Re-run the five Persian datasets with a documented, stratified train/validation/test split and check whether any training example also appears in the test set; if training and test texts overlap, the 96% accuracy would drop when the overlap is removed.

Watch

Extended reading notes

Core claim

The central claim is that cross-lingual transfer plus few-shot fine-tuning is enough for effective Persian sentiment classification, and that when domains are introduced incrementally, regularization is what prevents forgetting from erasing that gain. Working from 1 to 20 labeled examples per class across five Persian-domain datasets, the paper reports that mDeBERTa and XLM-RoBERTa reach roughly 96% accuracy and F1-score at 20 shots. In the incremental condition, knowledge distillation is the strongest strategy in extreme 1-shot regimes and rehearsal is consistently strong across shot counts, while Elastic Weight Consolidation is less stable and unregularized incremental training clearly suffers in low-shot settings. Joint fine-tuning without incremental steps also reaches the top accuracy at 20 shots, so the paper's precise claim is that well-regularized incremental adaptation is competitive with, and sometimes better than, training on all domains at once, not that incremental steps are always required.

Load-bearing premise

The reported 96% figures assume that the few-shot training examples are disjoint from the examples used for evaluation; the paper does not describe the train/validation/test split, so if the same texts appear on both sides the numbers would measure memorization rather than generalization.

Editorial extensions

If this is right

If the results hold, a usable Persian sentiment classifier can be built from about 20 labeled examples per class, a fraction of the data normally required.
In incremental pipelines, rehearsal and knowledge distillation are the safety mechanisms that matter; unregularized updates lose much of the cross-lingual knowledge.
At higher shot counts, simply fine-tuning on all domains at once is competitive or better, so incremental training is a strategy for constrained or streaming data rather than a universal upgrade.
The consistent gap between XLM-R and mDeBERTa on one side and DistilBERT on the other indicates that broad multilingual pre-training is the decisive ingredient in low-shot transfer.
The same recipe should generalize to other low-resource languages and text classification tasks, since nothing in the method is Persian-specific beyond the datasets.

Reading between the lines

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

Because the train/test split is unspecified in the paper, the headline numbers should be read as upper bounds until replication with a public split confirms them.
A testable extension is to reverse the dataset order (long complex reviews first, short tweets last); if performance drops, the paper's curriculum ordering is doing real work.
A comparison the paper does not include, fine-tuning each domain only on its own few-shot examples without cross-domain incremental training, would isolate whether cross-domain transfer or regularization is responsible for the gains.
The same pipeline could be probed zero-shot on a sixth Persian domain, such as news comments, to see whether incremental adaptation improves generalization or mainly memorizes the five training domains.

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.

Desk Editor's Note

read the letter

Quick take: this is a workmanlike empirical study, not a breakthrough. The useful part is the matrix: three multilingual models (XLM-R, mDeBERTa, DistilBERT) times six shot counts times five training regimes (no incremental, incremental with no regularization, EWC, distillation, rehearsal) over five Persian domain datasets. That is a lot of coverage, and the tables are readable. If the numbers are honest, they give a practical recipe: mDeBERTa and XLM-R with a few dozen examples per class can get to mid-90s accuracy on these domains.

But the central claim in the abstract and conclusion—that incremental learning improves performance—does not survive contact with the paper's own numbers. The stress-test is right. At 10, 15, and 20 shots, the No Incremental Learning baseline wins or ties in most of the model-by-shot comparisons. For instance, XLM-R at 20 shots: 96.62 F1 without incremental, 94.43 with rehearsal, 93.25 with distillation. mDeBERTa is similar. The incremental variants only clearly win at 1-shot. So the unqualified sentence "The results indicate that incremental learning methods can improve model performance" is simply false for the higher-shot settings. The paper even contains a more careful sentence in Section IV.B.4 admitting that No Incremental Learning provided the highest performance, which makes the abstract/conclusion overclaim look like an oversight rather than a deliberate misreading, but it is still a fundamental mismatch with their own evidence.

The methodological reporting is also thin. There is no train/validation/test split description, no random seeds, no multiple runs, no standard deviations. With few-shot data, a one or two point difference between methods is plausibly noise, and a leaky split could inflate the headline 96% figures. The paper does not describe how classes were balanced across splits, whether the same examples could appear in training and test, or even how the few-shot samples were selected. That makes the numbers underdetermined. The models used are already fine-tuned on multilingual sentiment datasets, so calling this "cross-lingual few-shot learning" is a bit generous—it is really continued fine-tuning on Persian, which is fine, but the framing overstates the novelty.

What is genuinely useful: the domain coverage (X, Instagram, Digikala, Snappfood, Taaghche) is the kind of diverse evaluation that low-resource sentiment analysis needs, and the comparison of EWC versus rehearsal versus distillation in a few-shot incremental setup is a reasonable empirical question. The paper deserves a serious referee because the experimental matrix is substantial and the flaws are fixable: re-analyze the tables, fix the overclaim, and release split details and code. But as it stands, I would not cite it for the incremental learning conclusion.

Recommendation: send it to peer review with the expectation of major revision. The editor should not desk reject it, but it needs the authors to either re-frame the contribution around the baseline's success or provide a rigorous demonstration of where incremental learning genuinely helps.

Referee Report

4 major / 5 minor

Summary. The paper studies few-shot cross-lingual sentiment analysis for Persian by fine-tuning three multilingual transformer models (XLM-RoBERTa, mDeBERTa, DistilBERT) on small numbers of examples from five Persian datasets (X, Instagram, Snappfood, Taaghche, Digikala). It compares incremental learning variants (no regularization, EWC, knowledge distillation, rehearsal) against a no-incremental-learning baseline, reporting accuracy and F1-score up to 96% and concluding that combining few-shot and incremental learning is effective, particularly with regularization.

Significance. If the experimental results were reproducible, the paper would be a useful empirical study of low-resource Persian sentiment analysis and of regularization techniques in domain-incremental learning. Its strengths include the use of five diverse Persian datasets, three model families, and three regularization baselines, as well as the use of publicly available pre-trained checkpoints. However, the paper's own results undermine its main claim about incremental learning, and the missing experimental protocol prevents the headline numbers from being verified. The significance currently rests on an overgeneralized conclusion rather than on the observed data.

major comments (4)

[Section V, Tables I and II] The conclusion that 'incremental learning methods can improve model performance' is not supported by the reported data. In most 10-, 15-, and 20-shot comparisons, the No Incremental Learning baseline achieves the highest F1-score: for example, XLM-R at 20 shots gives No Incremental 96.62 vs Rehearsal 94.43 and Knowledge Distillation 93.25; mDeBERTa at 20 shots gives 96.68 vs 93.37 and 93.37; DistilBERT at 20 shots gives 57.80 vs 52.72 and 53.98. Only at 1-shot do Knowledge Distillation (and sometimes Rehearsal) clearly beat the baseline, e.g., XLM-R 1-shot KD 90.95 vs No Inc 87.46 and mDeBERTa 1-shot KD 90.92 vs 87.47. Section IV.B.4 itself states that 'No Incremental Learning provided the highest performance.' The conclusion should be restricted to the extreme few-shot regime or supported by statistical tests.
[Section IV.A] The experimental setup does not describe the train/validation/test split, the number of random seeds or runs, standard deviations, or the exact procedure for drawing few-shot samples (e.g., random vs stratified, sampled per dataset before or after the incremental order). Without this information, the 96% accuracy figures in Tables I and II cannot be checked for data leakage, and the reported differences among methods might be within run-to-run noise. The authors must specify the split sizes, the number of test examples per dataset, whether training and test sets are disjoint, and report mean and standard deviation over multiple runs.
[Section IV.B and Tables I/II] The narrative text and the tables contain numerical inconsistencies. Section IV.B.1 reports XLM-R 1-shot Knowledge Distillation F1 of 91.32, but Table II shows 90.95; Section IV.B.2 reports mDeBERTa 1-shot Knowledge Distillation F1 of 91.19, but Table II shows 90.92 (91.19 is the accuracy in Table I); and Section IV.B.2 states that at 15 shots Rehearsal reached 94.01% F1 while No Incremental Learning had 87.66%, which matches no row of Table II (the mDeBERTa 15-shot values are 95.58 and 96.65, respectively). All reported values and their labels should be corrected and cross-checked against the tables.
[Tables I and II] The F1-score is not defined, and some F1 values are implausibly large relative to accuracy for the same configuration. For example, DistilBERT at 5 shots with No regularization has accuracy 52.24 in Table I but F1 of 78.60 in Table II. This large gap can arise from macro-averaging or positive-class F1 under class imbalance, but without an explicit definition and class-wise breakdown the metric cannot be interpreted. The authors should state whether F1 is macro, micro, or per-class and explain the accuracy-F1 discrepancies.

minor comments (5)

[Abstract and Introduction] The phrasing 'effectiveness of combining few-shot learning and incremental learning' overstates the results; the data only support a benefit in the 1-shot setting. Please rephrase to match the evidence.
[Section III.E] The claim that the curriculum order X, Instagram, Snappfood, Taaghche, Digikala was 'chosen based on the relative complexity of the texts' is not supported by any complexity measure. Either provide a quantitative criterion or describe this as an author choice.
[Section III.A] The statement that XLM-RoBERTa 'achieved approximately 23% higher accuracy than mBERT' cites reference [4] without a specific result. Please add a precise comparison or qualify the claim.
[Section IV.A] The term 'maximum of 10 training iterations per phase' is ambiguous; please specify whether this means epochs, optimization steps, or something else.
[References] Reference [18] contains the Persian title as 'Taaghche | .طاقچه' with an odd separator; also fix the hyphen in 'Huggingface' and the spacing in 'over fitting' (Section III).

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning found; the paper reports empirical measurements from external models and datasets, and its central claims do not reduce to their inputs by construction.

full rationale

The paper contains no derivation chain in which an output is defined in terms of an input. The three models are pre-trained externally on multilingual and sentiment corpora (Refs. [5], [7], [9], and [11]) and are then fine-tuned on Persian datasets (Refs. [14]-[18]); the reported accuracy and F1 values are measured outcomes of that procedure, not fitted quantities renamed as predictions. No parameter is fitted to a subset and then used to predict a quantity that equals the fit by construction, and no load-bearing premise is justified solely by a self-citation: the cited prior work is by third-party authors or external model and dataset releases. The abstract's claim that incremental learning improves performance is contradicted in the paper's own Tables I and II, for instance the No Incremental Learning baseline is highest in most 10-20 shot comparisons, but that is an internal consistency and evidence problem, not circularity. The undocumented train/validation/test partitioning in Section IV.A creates a potential leakage risk, but leakage is data contamination and does not make the argument circular. Accordingly, no circular step can be quoted with a specific reduction, and the honest finding is a score of 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim depends on standard transfer-learning assumptions, an undocumented test split, a hand-chosen curriculum order, and two hyperparameters. No new entities are introduced.

free parameters (2)

early_stopping_patience = 3
Chosen by hand in Section IV.A; stops training if validation loss does not improve for 3 consecutive periods, affecting all reported results.
max_iterations_per_phase = 10
Chosen by hand in Section IV.A; caps the number of training iterations per incremental phase, which may limit learning in low-shot settings.

assumptions (4)

domain assumption The few-shot training examples and the evaluation examples are disjoint, i.e., the reported test accuracy reflects generalization to unseen instances.
Section IV.A describes the incremental phases and validation-based early stopping but never specifies the train/validation/test split. All conclusions in Tables I and II assume such a disjoint split.
domain assumption Sentiment labels for the rating-based datasets (Digikala and Taaghche) are converted from numeric ratings using some consistent threshold; the threshold is not reported.
Section III-D states sentiment is inferred from rating without giving the mapping, so label quality depends on an undocumented assumption.
ad hoc to paper The curriculum order X, Instagram, Snappfood, Taaghche, Digikala, chosen by text complexity, is beneficial for incremental learning.
Section IV.A states the order is 'chosen based on the relative complexity of the texts,' an a priori judgment rather than a measured result; results may depend on this order.
domain assumption Cross-lingual transfer from multilingual pre-training is effective for Persian with few examples, i.e., the base models carry enough linguistic knowledge to adapt to Persian.
The entire methodology relies on the standard assumption that fine-tuning a multilingual model on a small Persian sample transfers the English/other-language abilities to Persian; this is the premise of the study.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cross-lingual Few-shot Learning for Persian Sentiment Analysis with Incremental Adaptation." pith.science (2026). https://pith.science/paper/MCG7PKMZ

@misc{pith2026250711634,
  author       = {Pith},
  title        = {Pith review of: Cross-lingual Few-shot Learning for Persian Sentiment Analysis with Incremental Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MCG7PKMZ}},
  note         = {Machine review of arXiv:2507.11634}
}

read the original abstract

This research examines cross-lingual sentiment analysis using few-shot learning and incremental learning methods in Persian. The main objective is to develop a model capable of performing sentiment analysis in Persian using limited data, while getting prior knowledge from high-resource languages. To achieve this, three pre-trained multilingual models (XLM-RoBERTa, mDeBERTa, and DistilBERT) were employed, which were fine-tuned using few-shot and incremental learning approaches on small samples of Persian data from diverse sources, including X, Instagram, Digikala, Snappfood, and Taaghche. This variety enabled the models to learn from a broad range of contexts. Experimental results show that the mDeBERTa and XLM-RoBERTa achieved high performances, reaching 96% accuracy on Persian sentiment analysis. These findings highlight the effectiveness of combining few-shot learning and incremental learning with multilingual pre-trained models.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 16 canonical work pages

[1]

Mozafari, M., Farahbakhsh, R., & Crespi, N. (2022). Cross -lingual few-shot hate speech and offensive language detection using meta learning. IEEE Access, 10, 14880-14896

work page 2022
[2]

Andrenšek, L., Koloski, B., Pelicon, A., Lavrač, N., Pollak, S., & Purver, M. (2024). Evaluating and explaining training strategies for zero-shot cross -lingual news sentiment analysis. arXiv preprint arXiv:2409.20054

work page Pith review arXiv 2024
[3]

Capuano, N., Greco, L., Ritrovato, P., & Vento, M. (2021). Sentiment analysis for customer relationship management: an incremental learning approach. Applied intelligence, 51, 3339-3352

work page 2021
[4]

& Stoyanov, V

Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., ... & Stoyanov, V. (2019). Unsupervised cross - lingual representation learning at scale. arXiv preprint arXiv:1911.02116

arXiv 2019
[5]

(2023, May 23)

terrencewee12/xlm-roberta-base-sentiment-multilingual-finetuned- v3 · Hugging Face. (2023, May 23). Huggingface.co. https://huggingface.co/terrencewee12/xlm-roberta-base-sentiment- multilingualfinetuned-v3

work page 2023
[6]

Argilla. (2020). twitter -coronavirus. Huggingface.co. https://huggingface.co/datasets/argilla/twittercoronavirus

work page 2020
[7]

Qiang, T. Y. (2023, June 24). multilingual -sentiments. Huggingface.co. https://huggingface.co/datasets/tyqiangz/multilingual-sentiments

work page 2023
[8]

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

arXiv 2019

Show all 21 references

[9]

(2023, May 23)

lxyuan/distilbert-base-multilingual-cased-sentiments-student · Hugging Face. (2023, May 23). Huggingface.co. https://huggingface.co/lxyuan/distilbert-base-multilingual-cased- sentiments-student

2023
[10]

He, P., Liu, X., Gao, J., & Chen, W. (2020). Deberta: Decoding - enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654

2020 arXiv
[11]

(2024, January 18)

sileod/mdeberta-v3-base-tasksource-nli · Hugging Face. (2024, January 18). Huggingface.co. https://huggingface.co/sileod/mdeberta-v3-base-tasksource-nli

2024
[12]

Yang, Y., Zhang, Y., Tar, C., & Baldridge, J. (2019). PAWS -X: A cross-lingual adversarial dataset for paraphrase identification. arXiv preprint arXiv:1908.11828

2019 arXiv
[13]

R., Schwenk, H., & Stoyanov, V

Conneau, A., Lample, G., Rinott, R., Williams, A., Bowman, S. R., Schwenk, H., & Stoyanov, V. (2018). XNLI: Evaluating cross - lingual sentence representations. arXiv preprint arXiv:1809.05053

2018 arXiv
[14]

MohammadAli MohammadKhani. (2023). Persian Twitter Dataset - Sentiment Analysis. Kaggle.com. https://www.kaggle.com/datasets/mohammadalimkh/persian- twitter-dataset-sentiment-analysis

2023
[15]

InstaText. (2020). Persian sentiment analysis dataset. Kaggle.com. https://www.kaggle.com/datasets/instatext/persian-sentiment- analysis-dataset

2020
[16]

Soheil Tehranipour. (2019). Digikala Comments (Persian Sentiment Analysis). Kaggle.com. https://www.kaggle.com/datasets/soheiltehranipour/digikala- comments-persian-sentiment-analysis

2019
[17]

Soheil Tehranipour. (2022). Snappfood - Persian Sentiment Analysis. Kaggle.com. https://www.kaggle.com/datasets/soheiltehranipour/snappfood- persian-sentiment-analysis

2022
[18]

Saeed Taghipour. (2019). Taaghche | .طاقچهKaggle.com. https://www.kaggle.com/datasets/saeedtqp/taaghche

2019
[19]

Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., ... & Hadsell, R. (2017). Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13), 3521-3526

2017
[20]

Robins, A. (1995). Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2), 123-146

1995
[21]

Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.0253

2015

Pith tools

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

[1] [1]

Mozafari, M., Farahbakhsh, R., & Crespi, N. (2022). Cross -lingual few-shot hate speech and offensive language detection using meta learning. IEEE Access, 10, 14880-14896

work page 2022

[2] [2]

Andrenšek, L., Koloski, B., Pelicon, A., Lavrač, N., Pollak, S., & Purver, M. (2024). Evaluating and explaining training strategies for zero-shot cross -lingual news sentiment analysis. arXiv preprint arXiv:2409.20054

work page Pith review arXiv 2024

[3] [3]

Capuano, N., Greco, L., Ritrovato, P., & Vento, M. (2021). Sentiment analysis for customer relationship management: an incremental learning approach. Applied intelligence, 51, 3339-3352

work page 2021

[4] [4]

& Stoyanov, V

Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., ... & Stoyanov, V. (2019). Unsupervised cross - lingual representation learning at scale. arXiv preprint arXiv:1911.02116

arXiv 2019

[5] [5]

(2023, May 23)

terrencewee12/xlm-roberta-base-sentiment-multilingual-finetuned- v3 · Hugging Face. (2023, May 23). Huggingface.co. https://huggingface.co/terrencewee12/xlm-roberta-base-sentiment- multilingualfinetuned-v3

work page 2023

[6] [6]

Argilla. (2020). twitter -coronavirus. Huggingface.co. https://huggingface.co/datasets/argilla/twittercoronavirus

work page 2020

[7] [7]

Qiang, T. Y. (2023, June 24). multilingual -sentiments. Huggingface.co. https://huggingface.co/datasets/tyqiangz/multilingual-sentiments

work page 2023

[8] [8]

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

arXiv 2019