Enhancing False-Sentence Pairs of BERT-Pair for Low-Frequency Aspect Category Detection

Aspect category detection (ACD), a task for detecting the aspect categories from texts, helps to extract the information customers need from review sentences automatically. BERT-pair was proposed, effectively solving this task by constructing auxiliary sentences from categories and classifying pairs of review and auxiliary sentences. In contrast, several product genres with multiple components include various low-frequency aspect categories in their reviews. Nevertheless, BERT-pair struggles to detect these categories; thus, the ACD performance for such products is poor. Moreover, to the best of our knowledge, no study on low-frequency category detection exists. Thus, we propose improved methods of BERT-pair to increase the performance for low-frequency categories. Our methods replace the review sentences in “false sentence pairs” that train BERT-pairs with similar sentences. This replacement enhances the diversity of sentences for low-frequency categories without additional training costs. Our experiment demonstrates that one of our methods increases the macro-F1 value for low-frequency categories by up to 0.11 points while maintaining the macro-F1 value for all categories.