ViCorpReviews: A Benchmark Dataset for Multi-Dimensional Sentiment and Hate Speech Detection in Vietnamese Workplace Context
Abstract
Online company review platforms have gained significant importance as they provide transparent insights into corporate culture and employee satisfaction. However, analyzing this feedback at scale remains challenging due to the linguistic complexity of workplace-specific discourse and the lack of high-quality, multi-dimensional annotated datasets, especially for low-resource languages like Vietnamese. This paper introduces ViCorpReviews, a novel benchmark dataset curated and annotated for two tasks: Aspect-based Sentiment Analysis (ABSA) and Hate Speech Detection (HSD). We implement a Human-in-the-loop annotation pipeline, leveraging the capabilities of Large Language Models (LLMs) to generate initial labels, which are subsequently refined and validated through a human verification process to ensure high-quality data. To establish robust baselines, we evaluate several well-known language models used for Vietnamese, specifically PhoBERT, XLM-RoBERTa, mT5, and ViT5. Our experimental results reveal the intricate patterns of workplace toxicity and sentiment, demonstrating that ABSA and HSD together provide complementary insights into the understanding of employee experience. The ViCorpReviews dataset serves as a foundational benchmark to foster future research in pattern recognition and natural language understanding within the corporate environment. Our dataset is released through this link: https://github.com/khanhlonguit/ViCorpReviews.