Bangla BERT for Hyperpartisan News Detection: A Semi-Supervised and Explainable AI Approach

Fatema Binte Hassan; Md Al Jubair; Md Masum Billah; Mohammad Mehadi Hasan; Sazzatul Yeakin; Zobayer Ahmed

REVIEW 4 major objections 5 minor 20 references

Bangla BERT for Hyperpartisan News Detection: A Semi-Supervised and Explainable AI Approach

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 A fine-tuned Bangla BERT model with semi-supervised pseudo-labeling detects hyperpartisan Bangla news at 95.65% accuracy, outperforming conventional machine learning classifiers.

desk verdict The 95.65% accuracy claim is contradicted by the paper's own confusion matrix, and the semi-supervised pipeline is too under-specified to trust; still worth a careful referee for the low-resource task itself. read the letter →

arxiv 2507.21242 v1 pith:CCW7N3YQ submitted 2025-07-28 cs.CL

Mohammad Mehadi Hasan , Fatema Binte Hassan , Md Al Jubair , Zobayer Ahmed , Sazzatul Yeakin , Md Masum Billah This is my paper

classification cs.CL

keywords hyperpartisannewsdetectionBanglaBERTsemi-supervisedlearningpseudo-labelingexplainableAILIMElow-resourceNLPpoliticalclassification

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

A fine-tuned Bangla BERT transformer, retrained with pseudo-labels from 47,987 unlabeled political articles, can detect hyperpartisan Bangla news more accurately than conventional machine learning classifiers. The paper reports 95.65% accuracy and a 95.44% F1-score on a held-out test split, and presents this as the first benchmark for Bangla hyperpartisan news detection. If the result holds, it matters because Bangla is a low-resource language where biased news spreads with few automated safeguards. The paper also uses LIME to show which words drive individual predictions, connecting the model's decisions to concrete partisan and neutral cues.

What carries the argument

The object that carries the argument is Bangla BERT, a transformer-based language model pre-trained on Bangla text and fine-tuned with a binary classification head. Around it sits a semi-supervised loop: the model first learns from a small labelled set, then assigns pseudo-labels to a much larger unlabeled political corpus, and is retrained on those predictions, with confidence-based filtering meant to keep low-quality labels out. The explainability component is LIME, a post-hoc method that fits a local interpretable model around a single prediction and highlights the word-level cues that pushed the article toward hyperpartisan or non-hyperpartisan.

What would settle it

Annotate a random sample of the unlabeled political articles by hand, compare those human labels with the model's pseudo-labels, and see whether agreement is high; if the pseudo-labels are no better than chance, the retraining loop cannot be the source of the reported accuracy. A second check is to train and test without any unlabeled data or augmentation and report whether the 95.65% score survives.

Watch

Extended reading notes

Core claim

The central claim is that fine-tuned Bangla BERT with semi-supervised learning outperforms four traditional classifiers on the binary task of hyperpartisan versus non-hyperpartisan Bangla political news. In the reported experiment, the model scores 95.65% accuracy, 95.65% precision, 95.24% recall, and a 95.44% F1-score, with 220 true positives, 242 true negatives, 20 false positives, and 11 false negatives. The authors describe the workflow as first training on a small augmented labelled set, then using the trained model to pseudo-label 47,987 unlabeled articles, and retraining on the combined data. The paper claims this is the first Bangla hyperpartisan detection study and the first use of LIME explainability in this setting, and it compares the F1-score against English and German systems from an earlier shared task.

Load-bearing premise

The pipeline assumes that the pseudo-labels Bangla BERT assigns to the 47,987 unlabeled articles are accurate enough to retrain on, and that the round-trip translated copies of the original labelled articles carry the same hyperpartisan labels as their sources.

Editorial extensions

If this is right

If the reported performance holds, Bangla BERT becomes the reference model that future Bangla hyperpartisan detectors will be compared against.
The margin over the strongest traditional model, 95.44% versus 93.55% F1, indicates that contextual language representations capture partisan cues that bag-of-word-style features miss.
The semi-supervised pipeline offers a template for low-resource languages with a small labelled corpus and a large pool of in-domain unlabeled articles.
The LIME explanations give media-monitoring and fact-checking systems a way to justify why a specific Bangla article was flagged as hyperpartisan.

Reading between the lines

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

Inference: the same pseudo-labelling recipe could transfer to other low-resource languages that have unlabeled news text, but the paper does not test cross-language transfer.
Inference: the round-trip translation step that grows the labelled set from 805 to 3,220 articles can be isolated by training on original-only and augmented data separately; the paper does not report that comparison.
Inference: the paper gives no confidence threshold or accuracy check for the pseudo-labels, so the contribution of the unlabeled retraining loop to the 95.65% score is not yet measured.
Inference: the cross-lingual comparison compares different datasets, so the F1 numbers are contextual, not a head-to-head win over English or German systems; an independently labelled Bangla benchmark would settle that.

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

The headline number is the problem. The abstract says 95.65% accuracy and 95.65% precision, but the confusion matrix in Figure 2 gives 220 TP, 11 FN, 242 TN, 20 FP. That is 462/493 = 93.7% accuracy and 220/240 = 91.7% precision, not what is reported. This is not a minor typo; it affects the central claim. The F1 of 95.44% also does not match the matrix. So the main result, as presented, is internally inconsistent and unreproducible.

What is new: this is the first attempt I know of to detect hyperpartisan news in Bangla using a Bangla-specific BERT with semi-supervised pseudo-labeling and LIME explanations. The motivation is legitimate—low-resource languages need work—and the setup is sensible: compare against TF-IDF baselines, use BNAD political articles, augment with round-trip translation. The LIME examples are illustrative, though not quantitatively evaluated.

The soft spots go beyond the arithmetic. The pseudo-labeling step (Section IV) is load-bearing: the model labels 47,987 unlabeled articles and then retrains on them. The paper gives no confidence threshold, no check on pseudo-label accuracy, and no statement that the unlabeled pool is disjoint from the test split. Without that, the 95.65% figure could come from the model seeing near-copies of test articles during the second training pass. Also, the claim that semi-supervised learning improves performance is unsupported because there is no ablation without the pseudo-labeling stage. The cross-lingual comparison against SemEval-2019 systems is not meaningful: different languages, different datasets, different tasks; F1 scores are not comparable across corpora. The annotation of the initial 805 labels is not described at all.

On the positive side, the related work is reasonably covered for a paper of this scope, and the baselines are standard. The writing is rough but understandable.

Who is this for? Someone working on Bangla NLP or content moderation in low-resource settings might get a starting point. But the current version should not be taken as a reliable benchmark.

Recommendation: send it to peer review, but with a clear expectation of major revision. The authors need to fix the confusion matrix arithmetic, specify the pseudo-labeling protocol, run an ablation, and release data/code. If those are addressed, the result might be a modest but useful contribution. As is, the central claim is unsupported.

Referee Report

4 major / 5 minor

Summary. The paper proposes fine-tuning Bangla BERT for hyperpartisan news detection in Bangla, using a small labeled set of political articles, round-trip translation augmentation, and a semi-supervised pseudo-labeling step over roughly 48,000 unlabeled articles from the BNAD corpus. It reports 95.65% accuracy, 95.65% precision, 95.24% recall, and 95.44% F1 for Bangla BERT, compares these against Logistic Regression, Random Forest, Naive Bayes, and SVM, and illustrates model decisions with LIME explanations. The authors claim this is the first Bangla hyperpartisan detection benchmark and that the model achieves state-of-the-art performance in a cross-lingual comparison.

Significance. If the reported performance were reliable, this would be a useful initial benchmark for Bangla hyperpartisan detection and a meaningful demonstration of semi-supervised transformer fine-tuning in a low-resource language. The use of a publicly available corpus and LIME-based explanations is a positive aspect of the study. However, the central empirical claim is undermined by an internal inconsistency between the reported metrics and the paper's own confusion matrix, and by the lack of details needed to evaluate the pseudo-labeling pipeline; the paper in its current form does not establish the claimed result.

major comments (4)

[Section V.A, Table III, Figure 2] The reported confusion matrix (220 TP, 11 FN, 242 TN, 20 FP; total 493) yields accuracy (220+242)/493 = 0.937, precision 220/(220+20) = 0.917, recall 220/(220+11) = 0.952, and F1 approximately 0.934, yet Table III and the abstract report 95.65% accuracy, 95.65% precision, and 95.44% F1. Only the recall value matches the table. The headline metrics therefore cannot be reproduced from the paper's own data, which directly undercuts the central claim that Bangla BERT outperforms the conventional models.
[Section IV, semi-supervised step] The pseudo-labeling stage is described in one sentence: 'the trained Bangla BERT model was used to label the unlabeled dataset and train again.' No confidence threshold, no number of accepted pseudo-labeled samples, no estimate of pseudo-label accuracy, and no statement that the unlabeled articles are disjoint from the 15% test split are provided. In addition, there is no ablation comparing the model with and without the semi-supervised stage. Consequently, the contribution of the semi-supervised loop to the reported accuracy cannot be assessed.
[Section III.C and III.D, dataset description] The data construction is internally inconsistent and incomplete. Round-trip translating 805 articles three times yields 3,220 samples, but Section III.D states a labeled set of 3,200 samples; the labeled (3,200) plus unlabeled (47,987) total 51,187, not the 48,831 extracted from BNAD; and the procedure by which the initial 805 articles were labeled is not described (no annotation criteria, annotator background, or agreement measure). These inconsistencies make the dataset construction impossible to reproduce or validate.
[Section V.C, Table IV] The cross-lingual benchmark comparison is not a valid basis for the 'state-of-the-art performance' claim. The SemEval-2019 systems cited in Table IV were evaluated on different languages, different datasets, and different label sources, so a direct comparison of F1 scores is not meaningful. The claim that a 95.44% F1 on the authors' private Bangla construction constitutes state-of-the-art performance is unsupported.

minor comments (5)

[Abstract and Index Terms] There are typos in the abstract and index terms: 'fine-tunized' should be 'fine-tuned,' and 'Hyper-plain' should be 'Hyperpartisan.'
[Section II and References] Several citations do not match their in-text attributions: the text credits Agerri for clustering-based semantic features, but reference [6] is listed as Isbister and Johansson; the German political-bias study cited as [12] is paired with a reference title about toxic language detection. These mismatches should be corrected.
[Figures 1-4] The manuscript refers to Figures 1, 2, 3, and 4, but the figures are not present in the text provided for review; if they were part of the submission, they must be embedded in the final version.
[Table II] Table II is not legible: the column layout is garbled, with hyperparameter names and values appearing in inconsistent cells. A clean table with one row per model and one row per hyperparameter would improve clarity.
[Tables III and IV] The tables mix decimal and percentage formats (e.g., 0.9565 versus 95.65%). The paper should use one consistent format and, ideally, report confidence intervals or significance tests for the difference between Bangla BERT and the best traditional models, which are numerically close.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported accuracy is an empirical measurement against human labels; the semi-supervised pseudo-labeling step is a validation risk, not a circular reduction.

full rationale

The central claim (95.65% accuracy, F1 95.44%) is an experimental result reported from a held-out test split, not a derivation from a premise that already contains the conclusion. The semi-supervised stage in Section IV ('the trained Bangla BERT model was used to label the unlabeled dataset and train again') does reuse the model's own outputs as pseudo-labels, but the final evaluation is on human-labeled data, so no fitted parameter or equation is defined in terms of the target metric by construction. The paper lacks a no-semi-supervised control experiment, so the claimed improvement from pseudo-labeling is not isolated; and the reported confusion matrix (220 TP, 11 FN, 242 TN, 20 FP; total 493) is arithmetically inconsistent with the stated precision and dataset totals. These are correctness and reproducibility concerns, not circularity. There are no load-bearing self-citations: the BNAD dataset citation [19] is external, and no uniqueness theorem or prior-work ansatz is invoked to force the Bangla BERT choice. The cross-lingual comparison is not circular, though the systems are not directly comparable across languages and label sources.

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

The central result depends on the quality of augmented and pseudo-labeled data, none of which is measured or released. The model itself and LIME are not new entities; the unstated adaptive choices are the pseudo-label threshold and augmentation multiplicity.

free parameters (3)

Pseudo-label confidence threshold
The threshold for accepting a pseudo-label from the unlabeled set is never specified, yet it controls the size and quality of the self-training set.
Round-trip translation multiplicity = 3
Three round-trip translations per labeled article are used without justification; changing this number changes the augmented training set.
Article truncation length = 500 tokens
Only the first 500 tokens are retained for BERT, chosen to fit the 512 token limit; this discards potentially relevant later content.

assumptions (4)

domain assumption The BNAD dataset provides correct article genres and that extracting political articles yields a coherent corpus.
Section III.A relies on BNAD's categories without an audit of labeling quality.
domain assumption Round-trip machine translation preserves the truth of the hyperpartisan label.
Section III.C uses GoogleTrans to create 3,220 samples from 805 labeled items; if translation shifts stance or sentiment, the new labels are wrong.
ad hoc to paper The model's own high-confidence predictions on unlabeled data are accurate enough to serve as training labels.
Section IV retrains on pseudo-labeled data without validation of pseudo-label quality.
domain assumption The 15% test split evaluates the fully retrained model without leakage from pseudo-labeled data.
The paper never specifies that unlabeled articles are disjoint from the test split.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bangla BERT for Hyperpartisan News Detection: A Semi-Supervised and Explainable AI Approach." pith.science (2026). https://pith.science/paper/CCW7N3YQ

@misc{pith2026250721242,
  author       = {Pith},
  title        = {Pith review of: Bangla BERT for Hyperpartisan News Detection: A Semi-Supervised and Explainable AI Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CCW7N3YQ}},
  note         = {Machine review of arXiv:2507.21242}
}

read the original abstract

In the current digital landscape, misinformation circulates rapidly, shaping public perception and causing societal divisions. It is difficult to identify hyperpartisan news in Bangla since there aren't many sophisticated natural language processing methods available for this low-resource language. Without effective detection methods, biased content can spread unchecked, posing serious risks to informed discourse. To address this gap, our research fine-tunes Bangla BERT. This is a state-of-the-art transformer-based model, designed to enhance classification accuracy for hyperpartisan news. We evaluate its performance against traditional machine learning models and implement semi-supervised learning to enhance predictions further. Not only that, we use LIME to provide transparent explanations of the model's decision-making process, which helps to build trust in its outcomes. With a remarkable accuracy score of 95.65%, Bangla BERT outperforms conventional approaches, according to our trial data. The findings of this study demonstrate the usefulness of transformer models even in environments with limited resources, which opens the door to further improvements in this area.

Figures

Figures reproduced from arXiv: 2507.21242 by the authors.

**Figure 1.** Data Preprocessing C. Data Augmentation The initial labeling of 805 articles was insufficient. To expand the dataset, we used data augmentation by translating Bangla text to English and back to Bangla three times using the GoogleTrans package. This process increased the data set to 3,220 samples. D. Dataset Description The dataset is composed of two parts: a labeled set with 3,200 samples (1,476 hyperpartisan and 1,… view at source ↗

**Figure 2.** Bangla BERT Confusion Matrix The confusion matrix analysis reveals particularly robust classification for positive instances, with 220 true positives and 11 false negative. For negative instances, the model shows good performance with 242 true negatives and 20 false positives. The overall F1-score of 0.9544 indicates balanced performance between precision and recall, suggesting that the model maintains consistency … view at source ↗

**Figure 4.** LIME explanation for a Hyperpartisan news text in Bangla [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 3.** Figure 3: LIME explanation for a Non-Hyperpartisan news text in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 19 canonical work pages

[1]

Semeval-2019 task 4: Hyper- partisan news detection,

J. Kiesel, M. Mestre, R. Shukla, E. Vincent, P. Adineh, D. Cor- ney, B. Stein, and M. Potthast, “Semeval-2019 task 4: Hyper- partisan news detection,” in Proceedings of the 13th Interna- tional Workshop on Semantic Evaluation , 2019, pp. 829–839

work page 2019
[2]

Bert: Pre-training of deep bidirectional transformers for language understanding,

J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

arXiv 2018
[3]

Shap: A unified approach to interpreting model predictions,

S. Lundberg and S.-i. Lee, “Shap: A unified approach to interpreting model predictions,” Advances in neural information processing systems, pp. 1–10, 2017

work page 2017
[4]

Detection of hyperpartisan news articles using natural language processing technique,

N. R. Naredla and F. F. Adedoyin, “Detection of hyperpartisan news articles using natural language processing technique,” In- ternational Journal of Information Management Data Insights , vol. 2, no. 1, p. 100064, 2022

work page 2022
[5]

Masking and transformer-based models for hyperpartisanship detection in news,

J. Sánchez-Junquera, P. Rosso, M. Montes, and S. P. Ponzetto, “Masking and transformer-based models for hyperpartisanship detection in news,” in Proceedings of the International Con- ference on Recent Advances in Natural Language Processing (RANLP 2021) , 2021, pp. 1244–1251

work page 2021
[6]

Dick-preston and morbo at semeval-2019 task 4: Transfer learning for hyperpartisan news detection,

T. Isbister and F. Johansson, “Dick-preston and morbo at semeval-2019 task 4: Transfer learning for hyperpartisan news detection,” in Proceedings of the 13th International Workshop on Semantic Evaluation , 2019, pp. 939–943

work page 2019
[7]

Harvey mudd college at semeval-2019 task 4: The dx beaumont hyper- partisan news detector,

E. Amason, J. Palanker, M. C. Shen, and J. Medero, “Harvey mudd college at semeval-2019 task 4: The dx beaumont hyper- partisan news detector,” in Proceedings of the 13th International Workshop on Semantic Evaluation , 2019, pp. 967–970

work page 2019
[8]

Team yeon-zi at semeval-2019 task 4: Hyperpartisan news detection by de-noising weakly-labeled data,

N. Lee, Z. Liu, and P. Fung, “Team yeon-zi at semeval-2019 task 4: Hyperpartisan news detection by de-noising weakly-labeled data,” in Proceedings of the 13th International Workshop on Semantic Evaluation , 2019, pp. 1052–1056

work page 2019

Show all 20 references

[9]

The sally smedley hyperpartisan news detector at semeval-2019 task 4,

K. Hanawa, S. Sasaki, H. Ouchi, J. Suzuki, and K. Inui, “The sally smedley hyperpartisan news detector at semeval-2019 task 4,” in Proceedings of the 13th International Workshop on Semantic Evaluation , 2019, pp. 1057–1061

2019
[10]

Hyperpartisanship, disinformation and political conversations on twitter: The brazilian presidential election of 2018,

R. Recuero, F. B. Soares, and A. Gruzd, “Hyperpartisanship, disinformation and political conversations on twitter: The brazilian presidential election of 2018,” in Proceedings of the in- ternational AAAI conference on Web and social media , vol. 14, 2020, pp. 569–578

2018
[11]

Computational assess- ment of hyperpartisanship in news titles,

H. Lyu, J. Pan, Z. Wang, and J. Luo, “Computational assess- ment of hyperpartisanship in news titles,” in Proceedings of the International AAAI Conference on Web and Social Media , vol. 18, 2024, pp. 999–1012

2024
[12]

Mitigating biases in toxic language detection through invariant rationalization,

Y.-S. Chuang, M. Gao, H. Luo, J. Glass, H.-y. Lee, Y.- N. Chen, and S.-W. Li, “Mitigating biases in toxic language detection through invariant rationalization,” arXiv preprint arXiv:2106.07240, 2021

2021 arXiv
[13]

Team xenophilius lovegood at semeval-2019 task 4: Hyperpar- tisanship classification using convolutional neural networks,

A. Zehe, L. Hettinger, S. Ernst, C. Hauptmann, and A. Hotho, “Team xenophilius lovegood at semeval-2019 task 4: Hyperpar- tisanship classification using convolutional neural networks,” in Proceedings of the 13th International Workshop on Semantic Evaluation, 2019, pp. 1047–1051

2019
[14]

Embedding-based approaches to hy- perpartisan news detection,

K. Mohan and P. Chen, “Embedding-based approaches to hy- perpartisan news detection,” arXiv preprint arXiv:2501.01370 , 2025

2025 arXiv
[15]

Hyperpartisan news and articles detection using bert and elmo,

G. K. W. Huang and J. C. Lee, “Hyperpartisan news and articles detection using bert and elmo,” in 2019 International Conference on Computer and Drone Applications (IConDA) . IEEE, 2019, pp. 29–32

2019
[16]

Hyperpartisan news clas- sification with elmo and bias feature

G. KI WEI HUANG and J. C. Lee, “Hyperpartisan news clas- sification with elmo and bias feature. ” Journal of Information Science & Engineering , vol. 37, no. 5, 2021

2021
[17]

UBC- NLP at SemEval-2019 task 4: Hyperpartisan news detection with attention-based Bi-LSTMs,

C. Zhang, A. Rajendran, and M. Abdul-Mageed, “UBC- NLP at SemEval-2019 task 4: Hyperpartisan news detection with attention-based Bi-LSTMs,” in Proceedings of the 13th International Workshop on Semantic Evaluation , J. May, E. Shutova, A. Herbelot, X. Zhu, M. Apidianaki, and S....

2019
[18]

Spider-jerusalem at semeval- 2019 task 4: hyperpartisan news detection,

A. Alabdulkarim and T. Alhindi, “Spider-jerusalem at semeval- 2019 task 4: hyperpartisan news detection,” in Proceedings of the 13th International Workshop on Semantic Evaluation , 2019, pp. 985–989

2019
[19]

Bangla news article dataset,

A. M. Saad, U. N. Mahi, M. S. Salim, and S. I. Hossain, “Bangla news article dataset,” Data in Brief , vol. 57, p. 110874,
[2024]

A vailable: https://www.sciencedirect.com/sci- ence/article/pii/S2352340924008382

[Online]. A vailable: https://www.sciencedirect.com/sci- ence/article/pii/S2352340924008382

Pith tools

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

[1] [1]

Semeval-2019 task 4: Hyper- partisan news detection,

J. Kiesel, M. Mestre, R. Shukla, E. Vincent, P. Adineh, D. Cor- ney, B. Stein, and M. Potthast, “Semeval-2019 task 4: Hyper- partisan news detection,” in Proceedings of the 13th Interna- tional Workshop on Semantic Evaluation , 2019, pp. 829–839

work page 2019

[2] [2]

Bert: Pre-training of deep bidirectional transformers for language understanding,

J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

arXiv 2018

[3] [3]

Shap: A unified approach to interpreting model predictions,

S. Lundberg and S.-i. Lee, “Shap: A unified approach to interpreting model predictions,” Advances in neural information processing systems, pp. 1–10, 2017

work page 2017

[4] [4]

Detection of hyperpartisan news articles using natural language processing technique,

N. R. Naredla and F. F. Adedoyin, “Detection of hyperpartisan news articles using natural language processing technique,” In- ternational Journal of Information Management Data Insights , vol. 2, no. 1, p. 100064, 2022

work page 2022

[5] [5]

Masking and transformer-based models for hyperpartisanship detection in news,

J. Sánchez-Junquera, P. Rosso, M. Montes, and S. P. Ponzetto, “Masking and transformer-based models for hyperpartisanship detection in news,” in Proceedings of the International Con- ference on Recent Advances in Natural Language Processing (RANLP 2021) , 2021, pp. 1244–1251

work page 2021

[6] [6]

Dick-preston and morbo at semeval-2019 task 4: Transfer learning for hyperpartisan news detection,

T. Isbister and F. Johansson, “Dick-preston and morbo at semeval-2019 task 4: Transfer learning for hyperpartisan news detection,” in Proceedings of the 13th International Workshop on Semantic Evaluation , 2019, pp. 939–943

work page 2019

[7] [7]

Harvey mudd college at semeval-2019 task 4: The dx beaumont hyper- partisan news detector,

E. Amason, J. Palanker, M. C. Shen, and J. Medero, “Harvey mudd college at semeval-2019 task 4: The dx beaumont hyper- partisan news detector,” in Proceedings of the 13th International Workshop on Semantic Evaluation , 2019, pp. 967–970

work page 2019

[8] [8]

Team yeon-zi at semeval-2019 task 4: Hyperpartisan news detection by de-noising weakly-labeled data,

N. Lee, Z. Liu, and P. Fung, “Team yeon-zi at semeval-2019 task 4: Hyperpartisan news detection by de-noising weakly-labeled data,” in Proceedings of the 13th International Workshop on Semantic Evaluation , 2019, pp. 1052–1056

work page 2019