UrduAspectNet: Fusing Transformers and Dual GCN for Urdu Aspect-Based Sentiment Detection

Abstract

Urdu, characterized by its intricate morphological structure and linguistic nuances, presents distinct challenges in computational sentiment analysis. Addressing these, we introduce ”UrduAspectNet” – a dedicated model tailored for Aspect-Based Sentiment Analysis (ABSA) in Urdu. Central to our approach is a rigorous preprocessing phase. Leveraging the Stanza library, we extract Part-of-Speech (POS) tags and lemmas, ensuring Urdu’s linguistic intricacies are aptly represented. To probe the effectiveness of different embeddings, we trained our model using both mBERT and XLM-R embeddings, comparing their performances to identify the most effective representation for Urdu ABSA. Recognizing the nuanced inter-relationships between words, especially in Urdu’s flexible syntactic constructs, our model incorporates a dual Graph Convolutional Network (GCN) layer.Addressing the challenge of the absence of a dedicated Urdu ABSA dataset, we curated our own, collecting over 4,603 news headlines from various domains, such as politics, entertainment, business, and sports. These headlines, sourced from diverse news platforms, not only identify prevalent aspects but also pinpoints their sentiment polarities, categorized as positive, negative, or neutral. Despite the inherent complexities of Urdu, such as its colloquial expressions and idioms, ”UrduAspectNet” showcases remarkable efficacy. Initial comparisons between mBERT and XLM-R embeddings integrated with dual GCN provide valuable insights into their respective strengths in the context of Urdu ABSA. With broad applications spanning media analytics, business insights, and socio-cultural analysis, ”UrduAspectNet” is positioned as a pivotal benchmark in Urdu ABSA research.