سامانه همایش های دانشگاه آزاد اسلامی واحد نجف آباد

صفحه اصلی / سومین کنفرانس بین المللی مهندسی و علوم کامپیوتر و اولین کنفرانس ملی سامانه های هوشمند، کلان داده و رایانش پیشرفته

Domain Graph Smoothing for Robust Multi-Domain Sentiment Classification

نویسندگان :

Sanaz Gouran shourakchali¹ Kamran Layeghi² Faraein Aeini³

1- دانشگاه آزاد تهران شمال 2- دانشگاه آزاد تهران شمال 3- دانشگاه آزاد واحد ساری

کلمات کلیدی :

sentiment classification،graph smoothing،multi-domain NLP،domain shift،text preprocessing

چکیده :

Lexical variability and contextual inconsistency remain fundamental challenges for sentiment classification, even within large-scale benchmark datasets such as IMDB. This paper introduces a lightweight and domain-agnostic preprocessing module, termed the Domain Graph Smoothing Layer (DGSL), designed to stabilize token-level representations prior to classification. Unlike graph neural network architectures, DGSL applies a single-step graph-based smoothing operation derived from sentence-level word co-occurrence statistics, introducing a structural prior without additional trainable parameters. This mechanism reduces embedding noise and enhances local semantic coherence while maintaining computational efficiency. Experimental results obtained exclusively on the IMDB movie review dataset demonstrate that integrating DGSL consistently improves the performance of strong baseline classifiers. Specifically, DGSL increases classification accuracy from 84.0% to 87.0% when applied to BiLSTM and from 88.0% to 90.0% when combined with BERT. Correspondingly, macro-F1 scores improve by 4 percentage points for BiLSTM and 2 percentage points for BERT, alongside consistent gains in precision, recall, and ROC–AUC. These results indicate that simple graph-smoothing priors can substantially enhance representation stability and generalization in sentiment classification without relying on complex graph architectures or explicit domain adaptation strategies.