Improving Opinion Mining Through Automatic Prompt Construction

Opinion mining is a fundamental task in natural language processing. This paper focuses on extracting opinion structures: triplets representing an opinion, a part of text involving an opinion role, and a relation between opinion and role. We utilize the T5 generative transformer for this purpose. It also adopts a multi-task learning approach inspired by successful previous studies to enhance performance. Nevertheless, the success of generative models heavily relies on the prompts provided in the input, as prompts customize the task at hand. To eliminate the need for human-based prompt design and improve performance, we propose Automatic Prompt Construction, which involves fine-tuning. Our proposed method is fully compatible with multi-task learning, as we did so in our investigations. We run a comprehensive set of experiments on Multi-Perspective Question Answering (MPQA) 2.0, a commonly utilized benchmark dataset in this domain. We observe a considerable performance boost by combining automatic prompt construction with multi-task learning. Besides, we develop a new method that re-uses a model from one problem setting to improve another model in another setting as a Transfer Learning application. Our results on the MPQA represent a new state-of-the-art and provide clear directions for future work.