Soroush Jaberi
CV (PDF)
Soroush Jaberi

Write

jaberi.soroush@gmail.com

Karaj, Iran · UTC+3:30

Elsewhere

Work Research 2025 Preprint

HVB-Persian

A fine-tuned ParsBERT classifier fused with VADER lexicon scores reaches F1 of 0.88, 0.94 and 0.95 on Digikala, Snappfood and the combined set of about 84,000 Persian reviews.

Preprint on SSRN

The problem

Sentiment in Persian product reviews is harder to classify than it looks. Reviews are written informally, labeled data is scarce, and negation and sarcasm weaken models that rely on simple polarity cues. A transformer fine-tuned on reviews reads context well, but its decisions are hard to inspect. A lexicon is transparent and cheap, but blind to context.

The approach

HVB-Persian uses both. A ParsBERT classifier is fine-tuned on the reviews and produces class scores from the whole sentence. In parallel, VADER produces a polarity score from word-level rules. The two are combined by a weighted score fusion, and the weights are tuned on validation data rather than fixed by hand.

Fig. 1. The method in one picture: contextual scores from a fine-tuned transformer and a rule-based polarity score, fused with weights chosen on validation data. Source: paper abstract.

Pairing a learned representation with an explicit, rule-based polarity signal is meant to make the pipeline more reliable, and easier to interpret, than either part alone.

Data and results

The reviews come from two Persian platforms, Digikala and Snappfood. I prepared both and merged them into a combined dataset of about 84,000 reviews.

0.88

F1 on Digikala

0.94

F1 on Snappfood

0.95

F1 on the combined dataset

These are the scores reported in the preprint, which has the full evaluation. A preprint has not been peer-reviewed.

Where it came from

This paper sits on earlier, smaller experiments on the same question, each public:

  • Classical baselines. Naive Bayes, Linear SVC, SGD and Random Forest classifiers on TF-IDF features, compared on accuracy, precision, recall and F1. All four land between 0.857 and 0.862 accuracy.
  • Persian sentiment with FastText. A three-class classifier (positive, negative, neutral) for Persian text, with 86% accuracy.

Where it is going

The same question is the subject of my M.Sc. thesis, Improving sentiment classification of Persian text using transformer-based and lexicon-based methods, supervised by Dr. Amineh Amini with Dr. Hadi Saboohi as co-advisor. I expect to defend it in January 2027.