PerSoMed: A Large-Scale Balanced Dataset for Persian Social Media Text Classification

Ebrahim Ansari; Isun Chehreh

REVIEW 4 major objections 6 minor 14 references

PerSoMed: A Large-Scale Balanced Dataset for Persian Social Media Text Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A new 36,000-post balanced Persian social media dataset becomes the first large-scale resource of its kind, and a Persian-specific transformer beats multilingual models on it.

desk verdict Useful Persian dataset effort undermined by ChatGPT-generated data and missing reproducibility details — worth revising, not rejecting. read the letter →

arxiv 2602.19333 v2 pith:QW6T2VAC submitted 2026-02-22 cs.CL cs.IRcs.SI

Isun Chehreh , Ebrahim Ansari This is my paper

classification cs.CLcs.IRcs.SI

keywords PersianNLPsocialmediatextclassificationbalanceddatasetdataaugmentationChatGPTTookaBERTclassimbalancebenchmark

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 to build the first large-scale, balanced dataset for Persian social media text classification, covering nine categories with 4,000 posts each. To reach balance, the authors combined real collected posts with synthetic examples generated mainly by ChatGPT, plus lexical replacement, and then benchmarked a range of models. The strongest empirical finding is that TookaBERT-Large, a Persian-specific transformer, achieves the best F1-score of 0.9621, outperforming multilingual models such as XLM-RoBERTa and FaBERT. If the dataset is accepted as representative of real Persian social media, it would provide a new benchmark for a low-resource language and support applications like trend detection and user profiling. The pith, therefore, is as much about the resource itself as about the demonstration that Persian-specific pretraining pays off for this task.

What carries the argument

The central object is the balanced dataset itself, produced by a two-pronged balancing pipeline: undersampling via semantic redundancy removal (ParsBERT embeddings plus cosine similarity to keep the most diverse posts) and oversampling via hybrid augmentation (lexical replacement for about 10% of synthetic samples and ChatGPT few-shot prompting for the other 90%). This pipeline is what turns a skewed collection of 50,000 human posts into a clean, category-balanced 36,000-post corpus. The benchmark then relies on fine-tuned transformer models, with TookaBERT-Large as the largest Persian-specific model, to demonstrate that language-specific pretraining extracts the most from the balanced data.

What would settle it

Collect a held-out set of genuine, human-written Persian social media posts (with no ChatGPT augmentation or label involvement) across the same nine categories and evaluate TookaBERT-Large trained on the balanced dataset; if its F1 drops well below 0.9621 (e.g., under 0.90), the assumption that synthetic posts represent real text is falsified.

Watch

Extended reading notes

Core claim

The paper introduces PerSoMed, a dataset of 36,000 Persian social media posts with 4,000 samples in each of nine categories (Economic, Artistic, Sports, Political, Social, Health, Psychological, Historical, Science & Technology). Around 60,000 raw posts were collected from Persian platforms, cleaned, and labeled via ChatGPT few-shot prompting with human verification. To correct an initial class imbalance, overrepresented categories were undersampled by removing semantically redundant posts using ParsBERT embeddings and cosine similarity, while underrepresented categories were augmented, with roughly 90% of the augmented content generated by ChatGPT and 10% by synonym replacement. The paper e

Load-bearing premise

The central assumption is that ChatGPT-generated synthetic posts are representative of real Persian social media text, so that models trained and evaluated on the augmented balanced dataset will perform similarly on genuine social media.

Editorial extensions

If this is right

The dataset provides a standardized evaluation resource for Persian social media text classification, which previously lacked a large balanced corpus across multiple categories.
The reported advantage of TookaBERT-Large over multilingual models suggests that Persian-specific pretraining can yield measurable gains on downstream social media tasks, guiding model choice for practitioners.
The hybrid augmentation strategy, combining lexical replacement with generative prompting, offers a template for other low-resource languages facing class imbalance.
The ablation study, showing that balancing improves all models and that BiLSTM is most sensitive to imbalance, reinforces that data-balancing techniques are a necessary step before model comparison.
The dataset's public release enables downstream applications such as trend detection, social behavior modeling, and user classification for Persian-language platforms.

Reading between the lines

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

Because the same LLM family (ChatGPT) generated roughly 90% of the augmented texts and also produced the initial labels, the benchmark may partly measure a classifier's ability to separate AI-generated texts into AI-instructed categories rather than cleanly classifying authentic human social media text.
A strong test of the dataset's external validity would be to hold out a purely human-written, non-augmented set of posts and measure how much F1 drops; if the drop is large, the reported 0.9621 may not transfer to real-world Persian social media.
The acknowledged overlap between Social and Political categories suggests that a hierarchical or multi-label classification scheme would better capture the data's structure and could improve scores on those classes.
The dataset's concentration on 2025 content, noted by the authors as a limitation, implies that temporal drift could degrade performance, so periodic re-collection or domain adaptation may be needed for sustained usefulness.

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

You should know two things about PerSoMed. First, it is a plausible answer to a real gap: there is no large balanced Persian social media classification benchmark, and the authors have collected a substantial amount of real Persian posts and organized them into nine categories. Second, the way the dataset is built makes its benchmark numbers hard to trust, and the paper does not currently give reviewers the tools to check.

What is genuinely new: the corpus itself. They collected ~50k real posts, used ChatGPT with human review to label, then balanced by undersampling majority classes and augmenting minority classes. For example, Historical had 497 real posts, so ~3,500 of its 4,000 final samples are ChatGPT-generated. The benchmark covers a decent spread of models, from BiLSTM to TookaBERT, and the finding that Persian-specific models beat multilingual ones is plausible and worth testing.

The soft spots are substantial. Most importantly, the dataset link is not actually in the paper — it says “Dataset Link” as a placeholder. For a dataset paper, that alone is a rejection-level problem. On top of that, there is no train/validation/test split, no hyperparameters, and no standard deviation on any metric, so the results cannot be reproduced. The paper also reports no inter-annotator agreement or details on the human review, making it hard to assess label quality.

The deeper issue is the ChatGPT confound. The same model family generated 90% of the augmentation for minority classes and also did the initial labeling. Unsurprisingly, class-wise results show Historical—the class with the most synthetic data—at 0.99 F1, while Social and Political, which are largely real, sit at 0.88 and 0.92. This pattern suggests the model may be recognizing stylistic fingerprints of ChatGPT rather than classifying genuine Persian social media text. The ablation in Table 5 compares balanced vs unbalanced data, and the unbalanced condition is mostly real, so the improvement from balancing could simply mean the model learns to exploit easy synthetic-pattern cues. Without a held-out evaluation on entirely real posts, we cannot know whether the reported 0.9621 F1 reflects actual field performance.

Who benefits: anyone working on Persian NLP and anyone thinking about how synthetic data contaminates benchmarks. But right now the paper is not usable as a resource because the data isn't accessible and the numbers are suspect. It deserves a serious referee—major revision, not desk reject—because the underlying resource is valuable and the flaws are fixable. If the authors release the dataset, specify the splits, and provide a real-only evaluation, this could become a useful contribution. I would not cite it in its current form, but I'd read the revision.

Referee Report

4 major / 6 minor

Summary. The paper introduces PerSoMed, a claimed first large-scale balanced Persian social media text classification dataset with 36,000 posts across nine categories, constructed from roughly 60,000 raw posts. Labeling uses ChatGPT few-shot prompting with human review; class imbalance is addressed by undersampling large classes with semantic redundancy removal and by augmenting small classes with lexical replacement and, for roughly 90% of the augmented data, ChatGPT-generated posts. The authors benchmark BiLSTM, XLM-RoBERTa with LoRA/AdaLoRA, FaBERT, an SBERT-based architecture, TookaBERT-Base, and TookaBERT-Large. They report that TookaBERT-Large achieves the best performance (Precision 0.9622, Recall 0.9621, F1 0.9621) and interpret this as evidence for the benefit of Persian-specific pretraining. The paper also includes an ablation study comparing balanced versus unbalanced data and reports class-wise results for TookaBERT-Large.

Significance. If the dataset and evaluation were methodologically sound, this would be a useful resource for Persian NLP: it is large, balanced, publicly released, and covers nine socially relevant categories with standardized benchmark numbers from several model families. The authors also provide a systematic comparison of Persian-specific and multilingual transformers, and they make their dataset available, which supports reproducibility. However, the central empirical claims rest on an evaluation protocol that has not been shown to be valid for the intended real-world task. Because the strongest stated contribution is a new benchmark and a new state of the art on it, the current evidence is not sufficient to accept the benchmark's validity or the superiority claim for TookaBERT-Large.

major comments (4)

[§4, Tables 3–5] No train/validation/test split is specified anywhere in Section 4. The reader cannot determine whether the same posts used for training also appear in the reported metrics, whether augmented data appears in the test set, or whether stratified splitting by class was used. All metrics are also point estimates with no standard deviations or repeated-run information. The SOTA claim in the abstract and Section 5 is therefore not robustly supported. Please specify the split (e.g., 80/10/10, stratified), ensure the test set is disjoint from training and validation, and report mean±std over at least three random seeds.
[§3.1, §4.1] The benchmark's validity is threatened by the interaction between data generation and evaluation. Underrepresented classes, such as Historical with initially 497 real posts, are augmented by ChatGPT-generated text until they reach 4,000 samples, with roughly 90% of all augmentation coming from ChatGPT few-shot prompting. Since the same model family also produced the initial labels, a classifier may achieve high F1 by recognizing distributional cues of ChatGPT-generated text rather than by learning to classify genuine Persian social media posts. No validation of the synthetic data's representativeness is provided: there is no real-only held-out evaluation, no human evaluation of generated posts, and no comparison of synthetic versus real text distributions. Please add a real-only test set for all classes and report classification performance on that set separately, or otherwise demonstrat
[§4.2, Table 5] The ablation does not isolate the effect of synthetic augmentation. The balanced condition changes two variables simultaneously: class proportions and the presence of ChatGPT-generated content. The unbalanced condition contains the original real posts, whereas the balanced condition contains undersampled real posts plus synthetic posts. Thus the F1 improvement from 0.8621 to 0.9621 for TookaBERT-Large could be due to the introduction of synthetic data that is easier to classify, not to the balancing per se. Please include an additional condition with balanced classes using only real posts (e.g., by further undersampling or by collecting more real data) and an additional condition with augmented data but unbalanced class proportions, to disentangle these factors.
[§3.1, 'Dataset Link' and §4.1] The claim that Persian-specific pretraining is superior is confounded by the synthetic-data issue. TookaBERT-Large's advantage over XLM-RoBERTa and FaBERT may partly reflect the fact that the training and evaluation distribution is heavily synthetic. The conclusion in Section 5 that 'demonstrating the effectiveness of Persian-specific pretraining' is load-bearing for the paper's contribution but is not justified unless the model is evaluated on real, non-augmented Persian social media posts. At minimum, report model performance on the real-only subset of the test data.

minor comments (6)

[§3.1, Table 1] The hashtag strings in Table 1 are rendered as garbled characters in the PDF (e.g., 'History (Ծֈ)'). Please use a Unicode font that displays Persian text correctly so the table is readable.
[§3.1] The dataset link is given as the placeholder text 'Dataset Link'. Provide a working URL and, ideally, a DOI for the dataset.
[§3.2] The paper does not report hyperparameters for any model (learning rate, batch size, number of epochs, LoRA rank, etc.). Reporting these details is necessary for reproducibility, especially for a claimed benchmark.
[§3.1] The preprocessing removes emojis and mentions, but social media text classification often benefits from these signals. Please justify this decision or at least note its effect on model performance and on the representativeness of the dataset.
[§4.1] Table 4 reports class-wise scores only for TookaBERT-Large. Class-wise results for the other models would help the reader assess whether the gap is uniform or driven by particular classes, especially the synthetic-heavy classes.
[General] There are several textual inconsistencies, including duplicated references in the introduction (e.g., 'T aha et al. (2024)' appears twice with different citation formats) and inconsistent spelling of 'efficiency' as 'eﬀiciency'. A copyedit pass would improve the manuscript.

Circularity Check

1 steps flagged · score 6.0 of 10

ChatGPT-generated minority-class posts make gold labels coincide with generator prompt conditions, so the headline F1 is partly circular.

self definitional [Section 3.1 (Dataset construction; ChatGPT labeling and augmentation)]
"For labeling, the ChatGPT API (OpenAI (2025)) was utilized with a few-shot prompting... multiple examples per category helped guide the model to accurately assign labels. ... the majority of augmentation—about 90%—relied on ChatGPT’s few-shot prompting capabilities. Carefully crafted prompts, enriched with multiple category-specific examples, guided the model to generate high-quality, coherent, and contextually relevant synthetic posts."

For the ~90% of augmented minority-class samples, the gold label is not an independent measurement of real social-media content: each synthetic post is generated by prompting ChatGPT with the category name and examples, so the correct label is, by construction, the input condition of the generator. The same ChatGPT API also produced the initial labels for real posts (with later human review). Thus a classifier's F1 on these samples partly measures recovery of ChatGPT's prompt/category conventions rather than classification of independently observed Persian social media text. The paper reports no real-only held-out evaluation and no synthetic-vs-real distribution comparison, so the headline TookaBERT-Large F1=0.9621 is partially circular with respect to dataset construction.

full rationale

This is a dataset/benchmark paper rather than a formal derivation, so there is no equation-level circularity and no fitted parameter is later renamed as a prediction. The self-citations (Chehreh et al. 2024a,b; Khojasteh et al. 2020) appear only in background/related-work contexts and are not load-bearing. The one substantive circularity is in dataset construction: for the roughly 90% of augmentation done via ChatGPT few-shot prompting, the category prompt is the same information later used as the gold label, and the same API also produced the initial labels for real posts. Consequently, a large part of the evaluation signal measures whether a classifier recovers ChatGPT's category-prompt conventions rather than an independent property of Persian social media text. The Limitations section does not disclose a real-only evaluation, and the unbalanced ablation in Table 5 still contains the same synthetic/real mix, so it cannot disentangle the confound. Because the effect is partial (majority classes derive from real posts and human review was applied), the paper is not wholly circular, but the headline F1 is partly constructed.

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

The central claim depends on hand-set data-construction choices and on the unvalidated premise that LLM-generated text can stand in for real Persian social media posts. No new theoretical entities are introduced.

free parameters (5)

samples_per_class = 4000
Chosen by hand to create a balanced dataset; determines size (36,000 = 9 × 4,000).
minimum_word_count = 3
Posts shorter than three words were excluded to maintain content richness.
augmentation_mix = ~10% lexical replacement / ~90% ChatGPT few-shot
Chosen after back-translation failed; no ablation justifies this exact split.
semantic_redundancy_selection = top-4000 lowest cosine similarity
Overrepresented classes were reduced to 4,000 samples by ParsBERT cosine similarity; the specific threshold is not reported.
chatgpt_prompt_design = unspecified
No prompt text, temperature, or decoding parameters are given for labeling or generation, making the data construction irreproducible.

assumptions (5)

domain assumption ParsBERT sentence embeddings capture semantic redundancy in Persian sufficiently for cosine-similarity deduplication.
Invoked in Section 3.1 for undersampling; not validated on this data.
domain assumption ChatGPT few-shot annotations are accurate enough after human verification.
No inter-annotator agreement or disagreement statistics are reported; the human verification process is not specified.
domain assumption The nine predefined categories are adequate and mutually intelligible for Persian social media content, and a single label is appropriate for each post.
The authors themselves note ambiguity between Social and Political; no multi-label or hierarchy analysis is provided.
ad hoc to paper Synthetic posts generated by ChatGPT closely mirror real Persian social media text.
Central premise of the balancing strategy; not empirically validated against real post statistics or human evaluation.
standard math Standard transformer fine-tuning and PEFT methods (LoRA, AdaLoRA, XLM-R) perform as documented in prior literature.
These are standard background results, not re-verified here.

how reviews work

0 comments

Cite this review

Pith. "Pith review of PerSoMed: A Large-Scale Balanced Dataset for Persian Social Media Text Classification." pith.science (2026). https://pith.science/paper/QW6T2VAC

@misc{pith2026260219333,
  author       = {Pith},
  title        = {Pith review of: PerSoMed: A Large-Scale Balanced Dataset for Persian Social Media Text Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QW6T2VAC}},
  note         = {Machine review of arXiv:2602.19333}
}

read the original abstract

This research introduces the first large-scale, well-balanced Persian social media text classification dataset, specifically designed to address the lack of comprehensive resources in this domain. The dataset comprises 36,000 posts across nine categories (Economic, Artistic, Sports, Political, Social, Health, Psychological, Historical, and Science & Technology), each containing 4,000 samples to ensure balanced class distribution. Data collection involved 60,000 raw posts from various Persian social media platforms, followed by rigorous preprocessing and hybrid annotation combining ChatGPT-based few-shot prompting with human verification. To mitigate class imbalance, we employed undersampling with semantic redundancy removal and advanced data augmentation strategies integrating lexical replacement and generative prompting. We benchmarked several models, including BiLSTM, XLM-RoBERTa (with LoRA and AdaLoRA adaptations), FaBERT, SBERT-based architectures, and the Persian-specific TookaBERT (Base and Large). Experimental results show that transformer-based models consistently outperform traditional neural networks, with TookaBERT-Large achieving the best performance (Precision: 0.9622, Recall: 0.9621, F1- score: 0.9621). Class-wise evaluation further confirms robust performance across all categories, though social and political texts exhibited slightly lower scores due to inherent ambiguity. This research presents a new high-quality dataset and provides comprehensive evaluations of cutting-edge models, establishing a solid foundation for further developments in Persian NLP, including trend analysis, social behavior modeling, and user classification. The dataset is publicly available to support future research endeavors.

Figures

Figures reproduced from arXiv: 2602.19333 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 6 linked inside Pith

[1]

Introduction The rapid growth of online textual content, fu- eled by social media platforms, news outlets, and collaborative knowledge bases, has introduced major challenges in organizing and interpreting unstructured data ( T aha et al. (2024)). Among these sources, social media posts stand out for their brevity, informality, and fast-changing na- ture. ...

2024
[2]

The re- viewed studies collectively address the challenges of handling short, noisy, and domain-specific texts, particularly those found in social media contexts

Related Work In this part of the paper, we review recent develop- ments in text classification and Persian-language resources that are relevant to our work. The re- viewed studies collectively address the challenges of handling short, noisy, and domain-specific texts, particularly those found in social media contexts. arXiv:2602.19333v1 [cs.CL] 22 Feb 202...

arXiv 2026
[3]

Political

APPROACH This section begins with a detailed overview of the dataset created for this investigation. We will first outline the steps involved in constructing the dataset, including data collection, pre-processing, and labeling, which form the foundation for the subsequent analysis. In the following section, we review the implemented models, covering both ...

2025
[4]

First, we report the main results, highlighting the performance of models on the preprocessed dataset

Experiments This section presents the experimental findings and is divided into two parts. First, we report the main results, highlighting the performance of models on the preprocessed dataset. Second, we present the ablation study, analyzing the impact of undersampling and oversampling. 4.1. Results Model Precision Recall F1 BiLSTM 0.8823 0.8807 0.8813 S...
[5]

Social” and “Political

Conclusion In this study, we introduced the first large-scale, balanced dataset for Persian social media text classification, addressing a critical gap in the availability of high-quality resources for Persian- language NLP . By collecting, preprocessing, and carefully annotating 60,000 posts across nine di- verse categories, we constructed a robust datas...
[6]

No personally identifiable information was stored or published

Ethics Statement This work involves the collection of publicly avail- able Persian social media posts. No personally identifiable information was stored or published. Usernames, profile information, and URLs were removed during preprocessing. All data was pro- cessed in accordance with ethical research guide- lines for handling public online content
[7]

So- cial

Limitations While the dataset covers nine major categories, the boundaries between some classes (e.g., “So- cial” and “Political”) can be ambiguous, leading to lower classification scores in these areas. More- over, the dataset reflects recent online discourse and may not fully generalize to future linguistic or topical trends. In addition, due to the lim...
[8]

Bibliographical References Tom Bourgeade, Silvia Casola, Adel Mahmoud Wizani, Cristina Bosco, et al. 2024. Data aug- mentation through back-translation for stereo- types and irony detection. In CEUR WORK- SHOP PROCEEDINGS , volume 3878, pages 90–97. CEUR-WS. Martin Juan José Bucher and Marco Martini. 2024. Fine-tuned’small’llms (still) significantly outpe...

arXiv 2024

Show all 14 references

[74]

Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao

IEEE. Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao
[618]

Huan Ma, Changqing Zhang, Yatao Bian, Lemao Liu, Zhirui Zhang, Peilin Zhao, Shu Zhang, Huazhu Fu, Qinghua Hu, and Bingzhe Wu

Springer. Huan Ma, Changqing Zhang, Yatao Bian, Lemao Liu, Zhirui Zhang, Peilin Zhao, Shu Zhang, Huazhu Fu, Qinghua Hu, and Bingzhe Wu. 2023. Fairness-guided few-shot prompting for large language models. Advances in Neural Informa- tion Processing Systems, 36:43136–43155. Most...

2023
[2019]

arXiv preprint arXiv:1911.02116

Unsupervised cross-lingual represen- tation learning at scale. arXiv preprint arXiv:1911.02116. Claude Coulombe. 2018. Text data augmentation made simple by leveraging nlp cloud apis. arXiv preprint arXiv:1812.04718. Aleksandra Edwards and Jose Camacho- Collados. 2024. Languag...

1911 arXiv
[2021]

Neural Pro- cessing Letters, 53(6):3831–3847

Parsbert: Transformer-based model for persian language understanding. Neural Pro- cessing Letters, 53(6):3831–3847. Omar Galal, Ahmed H Abdel-Gawad, and Mona Farouk. 2024. Rethinking of bert sentence em- bedding for text classification. Neural Comput- ing and Applications , 36...

2024 arXiv
[2023]

arXiv preprint arXiv:2303.10512

Adalora: Adaptive budget allocation for parameter-eﬀicient fine-tuning. arXiv preprint arXiv:2303.10512. Yazhou Zhang, Mengyao Wang, Qiuchi Li, Prayag Tiwari, and Jing Qin. 2025. Pushing the limit of llm capacity for text classification. In Compan- ion Proceedings of the ACM o...

2025 arXiv
[2024]

arXiv preprint arXiv:2402.06617

Fabert: Pre-training bert on persian blogs. arXiv preprint arXiv:2402.06617 . OpenAI. 2025. Api reference. https: //platform.openai.com/docs/ api-reference. Accessed: 2025-08-12. Adamu Gaston Philipo, Doreen Sebastian Sar- watt, Jianguo Ding, Mahmoud Daneshmand, and Huansheng ...

2025 arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Introduction The rapid growth of online textual content, fu- eled by social media platforms, news outlets, and collaborative knowledge bases, has introduced major challenges in organizing and interpreting unstructured data ( T aha et al. (2024)). Among these sources, social media posts stand out for their brevity, informality, and fast-changing na- ture. ...

2024

[2] [2]

The re- viewed studies collectively address the challenges of handling short, noisy, and domain-specific texts, particularly those found in social media contexts

Related Work In this part of the paper, we review recent develop- ments in text classification and Persian-language resources that are relevant to our work. The re- viewed studies collectively address the challenges of handling short, noisy, and domain-specific texts, particularly those found in social media contexts. arXiv:2602.19333v1 [cs.CL] 22 Feb 202...

arXiv 2026

[3] [3]

Political

APPROACH This section begins with a detailed overview of the dataset created for this investigation. We will first outline the steps involved in constructing the dataset, including data collection, pre-processing, and labeling, which form the foundation for the subsequent analysis. In the following section, we review the implemented models, covering both ...

2025

[4] [4]

First, we report the main results, highlighting the performance of models on the preprocessed dataset

Experiments This section presents the experimental findings and is divided into two parts. First, we report the main results, highlighting the performance of models on the preprocessed dataset. Second, we present the ablation study, analyzing the impact of undersampling and oversampling. 4.1. Results Model Precision Recall F1 BiLSTM 0.8823 0.8807 0.8813 S...

[5] [5]

Social” and “Political

Conclusion In this study, we introduced the first large-scale, balanced dataset for Persian social media text classification, addressing a critical gap in the availability of high-quality resources for Persian- language NLP . By collecting, preprocessing, and carefully annotating 60,000 posts across nine di- verse categories, we constructed a robust datas...

[6] [6]

No personally identifiable information was stored or published

Ethics Statement This work involves the collection of publicly avail- able Persian social media posts. No personally identifiable information was stored or published. Usernames, profile information, and URLs were removed during preprocessing. All data was pro- cessed in accordance with ethical research guide- lines for handling public online content

[7] [7]

So- cial

Limitations While the dataset covers nine major categories, the boundaries between some classes (e.g., “So- cial” and “Political”) can be ambiguous, leading to lower classification scores in these areas. More- over, the dataset reflects recent online discourse and may not fully generalize to future linguistic or topical trends. In addition, due to the lim...

[8] [8]

Bibliographical References Tom Bourgeade, Silvia Casola, Adel Mahmoud Wizani, Cristina Bosco, et al. 2024. Data aug- mentation through back-translation for stereo- types and irony detection. In CEUR WORK- SHOP PROCEEDINGS , volume 3878, pages 90–97. CEUR-WS. Martin Juan José Bucher and Marco Martini. 2024. Fine-tuned’small’llms (still) significantly outpe...

arXiv 2024