Generative Opinion Triplet Extraction Using Pretrained Language Model

Abstract

Previous researches on opinion triplet extraction for Indonesian reviews in the hotel domain has been conducted in a discriminative manner using a sequence labeling approach. However, the opinion triplet extracted by those research is limited to explicit opinion triplet only, while neglecting the opinion triplets that contains implicit aspects. In this paper, we build a model that can perform opinion triplet extraction with explicit and implicit aspects based on the seq2seq approach. The system can extract opinion triplets using pre-trained language models that have been fine-tuned on datasets with label in extraction-style paradigm. The generated text then can be extracted to get the opinion triplets. By transforming opinion triplet extraction into a text generation problem with the help of a pre-trained language model (IndoT5), we are able to improve the F1 score by 4% when compared to the findings of earlier studies.