Multi-Label Aspect-Sentiment Classification on Indonesian Cosmetic Product Reviews with IndoBERT Model

Abstract

For an existing cosmetic company to expand, it is crucial to understand customers’ opinions regarding cosmetic products through product reviews. Aspect-based sentiment classification (ABSC), which consists of text representation and classification stages, is typically employed to automatically extract the interested insights from review. Existing studies of ABSC primarily used single-label classification, which fails to capture relationships between multiple aspects in a review. Additionally, the use of contextual embeddings like IndoBERT for representing Indonesian-language cosmetic product reviews has been underexplored. This study addresses these issues by developing a multi-label classification model that leverages IndoBERT, including IndoBERT[b], IndoBERT[k], and IndoBERTweet, to better represent context and capture relationships across multiple aspects in a review. The model is trained and evaluated using a dataset of Indonesian-language cosmetic product reviews from Female Daily. The multi-label models can be constructed using IndoBERT directly as end-to-end model or employing IndoBERT solely as word embedding model. The latter model, also known as conventional multi-label model, needs to be coupled with problem transformation approach and classifier for classification. Single label classification model with Word2Vec serves as baseline to assess the improvement of multi-label model’s performance on Female Daily cosmetic product reviews dataset. The empirical results revealed that the multi-label approach was more effective in identifying sentiments for pre-defined aspects in reviews. Among the models, end-to-end IndoBERT[b] achieved the highest accuracy (86.98%), while conventional multi-label models combining IndoBERT[b], Label Powerset (LP), and Support Vector Machine (SVM) performed best with 69.64%. This study is significant as it provides a more generalized understanding of the BERT embedding within the context of multi-labels classification and explores the effect of contextual embedding in the cosmetic domain.