Niloofar Ranjbar

Aspect-Based Sentiment Analysis (ABSA) is less developed for Persian than for English, particularly for extracting complete aspect–category–opinion–sentiment structures rather than classifying sentiment for predefined aspects. This paper presents a generative Persian ABSA framework based on fine-tuned LLaMA3-family instruction models. Given a Persian review, the model generates a JSON array of quadruples containing aspect terms, categories, opinion spans, and sentiment labels. The training data are drawn from a GPT-3.5-assisted Persian ABSA corpus whose annotations were reviewed and filtered by human annotators; we therefore treat it as GPT-assisted/human-validated rather than purely human-authored. To reduce circular evaluation, we also construct an independent human-only test set from raw Persian reviews. Three annotators manually annotated 240 reviews from scratch, and a two-of-three consensus procedure produced 555 gold aspect-level labels from 239 reviews. We evaluate zero-shot LLaMA3, fine-tuned LLaMA3 models, and Dorna-LLaMA3-8B-Instruct as a Persian-oriented generative baseline under the same schema, parser, and metrics. On the human-only test set, zero-shot LLaMA3 obtains 0.1426 aspect F1 and 0.1185 aspect–sentiment F1. Fine-tuned LLaMA3-CG improves the five-seed mean scores to 0.5477±0.0229 and 0.4394±0.0212, while Dorna-CG reaches 0.5809±0.0123 and 0.4897±0.0194. Adding ParsiNLU substantially improves in-domain ParsiNLU performance but reduces human-only performance, revealing a domain and annotation-style trade-off.