Joint Span-to-Span and LLM Training for Aspect Sentiment Triplet Extraction with Long Reviews

Most existing methods for Aspect Sentiment Triplet Extraction (ASTE) are primarily developed and evaluated on short reviews of 4 to 20 tokens. While effective in benchmark datasets, these models struggle in real-world scenarios like app reviews on the Google Play Store, where user feedback is often much longer. Token-length limitations cause important opinion phrases to be truncated or lost, reducing sentiment analysis accuracy. As a result, current ASTE models often fail to capture complete aspect-opinion-sentiment triplets in long reviews. This limits their practical usefulness in applications that require a full understanding of user sentiment.To address this issue, this study proposes a novel ASTE approach tailored for long reviews by jointly training a span-to-span model with a Large Language Model (LLM). The span-to-span model is first used to extract the most relevant aspects, which are then fed into an LLM using T5 to extract the corresponding opinions and sentiment polarities. The two models are trained simultaneously through an integrated loss function to optimize performance. Findings indicate that the T5-base demonstrates the highest F1-Score of 76.3%, surpassing previous methods in ASTE for long reviews.