Aspect-Based Sentiment Analysis of Multilingual Customer Product Reviews: Leveraging Unsupervised Aspect Extraction with ABAE for XLM-RoBERTa-based Classification

This study tackles the challenges of Aspect-Based Sentiment Analysis (ABSA) in noisy, multilingual customer reviews, using the Multilingual Amazon Reviews Corpus. We propose a two-stage process: first, an unsupervised Attention-based Aspect Extraction (ABAE) model identifies aspect labels. To enable cross-lingual aspect discovery, ABAE leverages pre-trained FastText word embeddings aligned into a common semantic space using MUSE. Second, a multilingual model: XLM-RoBERTa is fine-tuned to classify sentiment (Positive, Negative, or Neutral) for these extracted aspects within their review context. Evaluation is done using accuracy, precision, recall, and F1-score. Our experimental results show that while the extraction and sentiment classification can maintain a reasonable performance, the complexities of multilingual data introduce challenges that may affect overall model effectiveness.