Vi-AbSQA: Multi-task Prompt Instruction Tuning Model for Vietnamese Aspect-based Sentiment Quadruple Analysis

Abstract

Aspect-based sentiment analysis (ABSA) has recently received considerable attention within the Natural Language Processing (NLP) community, especially for complex tasks like triplet extraction or quadruplet prediction. However, most existing studies focus on high-resource languages. In this paper, we construct a challenging benchmark dataset for Vietnamese Aspect-based Sentiment Quadruple Analysis (AbSQA), where each sentence can contain explicit and implicit aspects and opinion terms. Moreover, each sample includes at least two aspect categories with different sentiments. We release this dataset for free research purposes, believing it will push forward research in this field. In addition, we present a generative-based approach to address the AbSQA task using a multitask instruction prompt tuning framework. Specifically, we design an effective generation paradigm that leverages instruction prompts to provide more information about the task. Besides, our model leverages relational information by designing separate sub-tasks based on the quadruplet elements and fine-tunes the transformer-based pretrained generative models in a multi-task manner. The experimental results demonstrate that our approach outperforms previously established extraction-based and generative-based methods, as well as the baseline variants.