Document Type : Original/Review Paper
Authors
Department of Computer Engineering, Na.C., Islamic Azad University, Najafabad, Iran
Abstract
The rapid growth of digital platforms has produced an unprecedented volume of user-generated reviews, making automated sentiment analysis increasingly important. Although transformer-based models have advanced sentiment classification, standard transformer classifiers do not receive dependency connectivity as a separate structural representation, a limitation that is especially relevant for morphologically rich languages such as Persian. This study introduces the Dependency-Based Importance Graph (DBIG), which converts dependency parses into undirected graphs, summarizes each word’s structural prominence with Degree Centrality and PageRank, and fuses the resulting vector with the transformer’s contextual representation. Unlike graph-based approaches that add trainable graph-learning layers, DBIG introduces no message-passing components and leaves the transformer unchanged. Experiments on a large-scale Persian Digikala review dataset for three-class sentiment classification (recommended, not recommended, no idea), spanning four transformer backbones, show that the effect of DBIG differs markedly across backbones. For mBERT and ParsBERT, DBIG yields large, statistically significant gains: ParsBERT+DBIG raises macro-F1 from 0.5709 to 0.7119 on the balanced dataset and accuracy from 0.6189 to 0.8012 on the imbalanced dataset. For AriaBERT, the augmentation produces no significant change on the balanced dataset but a small, statistically significant improvement on the imbalanced dataset. For FaBERT, it produces a small but statistically significant decline on both datasets. These findings indicate that the effect of dependency-derived structural augmentation varies across pretrained backbones, and differences in the syntactic and stylistic information encoded during pretraining are a plausible, though untested, explanation.
Keywords
- Sentiment Analysis
- Machine Learning algorithm
- Dependency-Based Importance Graph
- Word Vectorization
- Persian Transformer
Main Subjects
)