Publications
カンファレンス (国際) Generating Fine-Grained Aspect Names from Movie Review Sentences Using Generative Language Model
Tomohiro Ishii (Aoyama Gakuin univ.), Yoshiyuki Shoji (Aoyama Gakuin univ.), Takehiro Yamamoto (Hyougo univ.), Hiroaki Ohshima (Hyougo univ.), Sumio Fujita, Martin J. Dürst (Aoyama Gakuin univ.)
The 25th International Conference on Information Integration and Web Intelligence (iiWAS2023)
2023.11.22
This paper proposes a method for identifying an aspect highlighted in a sentence from a movie review, utilizing a generative language model. For example, the aspect “SFX Techniques” is identified for the sentence “The explosions in cosmic space were realistic.” Classically, aspects are commonly estimated in the field of opinion mining within product reviews with classification or extraction approaches. However, because the aspects of movie reviews are diverse and innumerable, they cannot be listed in advance. Thus, we propose a generation-based approach using a generative language model to identify the aspect of a review sentence. We adopt T5 (Text-to-Text Transfer Transformer), a modern generative language model, providing additional pre-training and fine-tuning to reduce the training data. To verify the effectiveness of the learning techniques thus adopted, we conducted an experiment incorporating reviews of Yahoo! movies. Manual labeling of the correctness and diversity of the aspect names generated shows that our method can generates a variety of fine-grained aspect names using little training data.
Paper :
Generating Fine-Grained Aspect Names from Movie Review Sentences Using Generative Language Model
（外部サイト）