A Fine-grained Sentiment Analysis Method Based On A Unified Framework

Aspect-Category-Opinion-Sentiment (ACOS) is a novel task in the field of fine-grained sentiment analysis, targeting the detection of implicit terms within aspect-based analysis. ACOS is designed with the objective of isolating sentiment quadruples within a specified textual sentence. The existing approach principally relies on scaling up Aspect-Opinion-Sentiment (AOS) triple extraction pipeline method. There will separate the ACOS method into multiple subtasks, likely leading to error propagation. To handle the above problems, we suggest a unified grid tagging framework for sentiment quadruple extraction. Our proposed framework follows the encoder-decoder manner with a core inference stage in the middle of encoding and decoding. And then, an encoder is assigned the initialization of the word vector, and the specific decoder is assigned the final quadruple extraction. During the inference phase, we additionally suggest a hierarchical correction strategy to correct the probabilities from different sub-classifiers. Our experimental results on the Restaurant-ACOS and Laptop-ACOS datasets demonstrate that proposed framework outperforms the baselines, achieving 7.75% and 3.2% improvement in F1-score, respectively.