📄 Identifying Political Bias in Arabic News Articles

Summary

The paper tackles automatic detection of political slant in Arabic journalism, using coverage of the Palestinian issue as its test case. Arabic is hard here because of rich morphology, dialect variety and few labelled resources. The authors scraped opinion articles from Al Jazeera Arabic and The Palestinian Information Centre (Palinfo), spanning periods before and after October 7, 2023. With help from an expert in Arabic political discourse, they hand-labelled a small balanced seed set as Pro, Against or Neutral. They then enlarged it with several text-augmentation techniques. A fine-tuned MARBERT supplied [CLS] embeddings, which fed four classifiers: Logistic Regression, SVM, Random Forest and LSTM. Spark handled the processing so the pipeline could later scale. The MARBERT + LSTM combination scored best, at 0.87 accuracy, and was used to label the rest of the unannotated collection. Neutral articles remained the hardest class to recognise. The work matters as one of the few Arabic-specific resources for media-bias classification, and it argues that domain-adapted transformers can pick up subtle rhetorical cues.

Key points

• Seed data and augmentation: 218 expert-annotated articles (73 Pro, 73 Against, 72 Neutral) were grown to 1,782 (670 Pro, 584 Neutral, 528 Against). The techniques were:
  • multi-step back-translation (argostranslate);
  • GPT-2 paraphrasing;
  • mBART summarization;
  • AraBERT-driven synonym substitution.
• Representation and split: MARBERT was fine-tuned, and its [CLS] vectors were stored in a DataFrame with numeric labels. A stratified train/test split was taken from the augmented set, and the evaluation portion has roughly 134 Pro / 117 Against / 106 Neutral items.
• Results:
  • LSTM reached the best accuracy, 0.87.
  • SVM and Logistic Regression landed around 0.80.
  • Random Forest reached about 0.82, but its results were less stable.
  • Logistic Regression's F1 was 0.82 on Pro and 0.75 on Against.
• Imbalance handling: SMOTE and class weighting were applied, and several vectorization schemes were compared to test feature robustness.
• Limitations:
  • Spark ran only in local mode, not on a real cluster.
  • Neutral consistently behaved like a minority, ill-defined class.
  • Planned future work includes Kafka streaming, more dialects and multimodal sources.

Interesting ideas

• I1. Label a tiny, carefully balanced expert seed set, train on it, and let the strongest model annotate the large unlabelled remainder; this trades expert hours for model-generated coverage.
• I2. Mix several unrelated augmentation routes (chained back-translation, paraphrase, summarization, synonym swap) so variants differ along different axes rather than one.
• I3. Keep one shared representation fixed and compare classifier families of very different kinds on it, which separates how much quality comes from the features versus the decision model.
• I4. The authors note two possible readings of the weak Neutral class: either neutral texts are genuinely rare, or the models cannot detect an absent signal. They treat "no stance" as a separate hypothesis worth testing.
• I5. Collecting data on both sides of a major event (October 7, 2023) gives temporal diversity, so the model is not tied to one news moment.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): the same bootstrapping could seed frame or speech-act labels from a few hand-checked texts, provided the silver labels are then cross-checked by independent witnesses before training.
• I2 (learn): such meaning-preserving variants can serve as invariance tests, where frames and world states extracted from a sentence and its rewrites should match, and mismatches flag fragile rules.
• I3 (learn): when evaluating a parser or frame extractor, freeze the upstream tree and swap only the downstream component, so gains are attributed to the right stage.
• I4 (frames): stance, irony and implicature modules should explicitly model a "no marked attitude" outcome with its own evidence, instead of defaulting to it when cues are missing.
• I5 (answer): test sets for QA readers could deliberately span texts from different periods or genres, to check that answers do not depend on period-specific vocabulary.