POS-ATAEPE-BiLSTM: an aspect-based sentiment analysis algorithm considering part-of-speech embedding

Aspect-based sentiment analysis (ABSA) is a granular sentiment classification task that involves identifying sentiment polarities toward aspects in a sentence. Performing ABSA on online e-commerce reviews is essential for understanding customers’ opinions about specific aspects of a product or service, which can help businesses make informed decisions to improve overall customer satisfaction. Although previous research on ABSA has demonstrated considerable achievements, ABSA remains challenging in Chinese. First, most previous approaches ignored the extraction of aspect multiword in Chinese ABSA tasks, leading to the omission of aspect words. Moreover, words with a specific part-of-speech (POS) in Chinese text may contain sentiment information. To effectively capture the information of sentiment words, POS information can be more extensively applied in the ABSA model. Herein, we propose a novel model named POS-ATAEPE-BiLSTM for ABSA in Chinese. The proposed model considers words and the corresponding POS information for training word vectors. In addition, the embedding layer contains text embedding, aspect embedding, and POS embedding. This study provides several key contributions. First, we propose a multiple-aspect-word extraction algorithm based on POS tagging and consider single aspect words and aspect multiwords for pre-training, reducing the omission of aspect words. Second, we introduce a novel framework for ABSA, which is based on POS-driven word embedding–training and contains POS information in the embedding layer to improve the performance of the ABSA model. We use three pre-training corpora (i.e., Chinese Wikipedia, online reviews, and news articles) for pre-training and demonstrate the robustness of our proposed model in Chinese ABSA tasks.