Separating Representation from Decision: A Controlled Comparison of Classifier Heads on Frozen ParsBERT Representations for Persian News Topic Classification

Automatic topic classification of news is difficult in Persian because of rich morphology, inconsistent orthography and topical overlap between categories. Pretrained transformers such as ParsBERT provide strong contextual representations, but it is unclear how much of the task is solved by the representation and how much by the classifier placed on top of it. This study holds a frozen ParsBERT encoder fixed and compares three heads with different inductive biases: logistic regression (LR), a multi-layer perceptron (MLP) and a convolutional network (CNN) over the token matrix. On a seven-class corpus of ۵۰۰ annotated news articles (Cohen’s κ = .۸۲) with a stratified held-out test set of ۱۰۰ articles, LR performed best (accuracy ۰.۷۹, ۹۵% CI [۰.۷۰, ۰.۸۶]; macro-F۱ ۰.۷۸), followed by the CNN (۰.۷۷; macro-F۱ ۰.۶۶) and the MLP (۰.۶۸; macro-F۱ ۰.۴۹); all exceeded a majority baseline (۰.۳۱). Exact McNemar tests on the paired predictions showed no difference between LR and the CNN (p = .۸۳); LR exceeded the MLP (p = .۰۳۵), but not after Holm correction. Repeated stratified ۵-fold cross-validation with five seeds confirmed the pattern (accuracy: LR ۰.۷۸ ± ۰.۰۳, CNN ۰.۷۶ ± ۰.۰۴, MLP ۰.۷۰ ± ۰.۰۴): LR outperformed the MLP (corrected resampled t-test, p = .۰۰۲) but not the CNN (p = .۳۸), class weighting reduced the MLP’s collapse on minority classes, and full fine-tuning of ParsBERT added only ۲.۵ accuracy points. Frozen ParsBERT representations thus make Persian news topics largely linearly separable, and added decision-layer capacity brings no benefit when labelled data are scarce and imbalanced.

کلیدواژه ها:

Persian Natural Language Processing ، News Topic Classification ، Parsbert ، Transfer learning ، Classifier Head ، logistic regression ، convolutional neural network ، Low-Resource Languages