Aspect-based Sentiment Analysis as Machine Reading Comprehension

Proceedings of the 29th International Conference on Computational Linguistics, pages 2461–2471 October 12–17, 2022. 2461 Aspect-based Sentiment Analysis as Machine Reading Comprehension Yifei Yang1,2, Hai Zhao1,2,∗ 1Department of Computer Science and Engineering, Shanghai Jiao Tong University 2MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University {yifeiyang}@sjtu.edu.cn, zhaohai@cs.sjtu.edu.cn Abstract Existing studies typically handle aspect-based sentiment analysis by stacking multiple neural modules, which inevitably result in severe error propagation. Instead, we propose a novel end- to-end framework, MRCOOL: MRC-PrOmpt mOdeL framework, where numerous senti- ment aspects are elicited by a machine reading comprehension (MRC) model and their corre- sponding sentiment polarities are classified in a prompt learning way. Experiments show that our end-to-end framework consistently yields promising results on widely-used benchmark datasets which significantly outperform exist- ing state-of-the-art models or achieve compara- ble performance.1 1 Introduction Compared with traditional sentence-level or document-level sentiment analysis tasks, aspect- based sentiment analysis (ABSA) requires finer grained analysis on the texts and extracting more detailed information (Liu, 2012; Pontiki et al., 2014a). ABSA contains many subtasks, such as aspect category detection, opinion term extraction (OE), etc. Aspect term extraction (AE) and aspect- level sentiment classification (SC) are two elemen- tal subtasks of ABSA. AE means extracting the sentiment aspects from a given plain sentence and SC implies recognizing the sentiment polarities of the given aspects in a sentence. Combining the above two subtasks, aspect term extraction and sen- timent classification (AESC) establishes the third fundamental subtask. The AE, SC and AESC for the sentence Excellent food, although the interior could use some help. are given in Figure 1. In general, the existing mainstream approaches can be roughly divided into two brands. The first ∗Corresponding author. This work was supported by Key Projects of National Natural Science Foundation of China (U1836222 and 61733011). 1https://github.com/yangyifei729/ MRC4absa •AE food • interior Excellent food, although the interior could use some help. •SC food-positive • interior-negative •Aspect Term Extraction and Sentiment Classification (food, positive) • (interior, negative) Figure 1: An example of AE, SC and AESC. AE needs to identify the aspect terms food and interior. When given these two terms, SC should recognize the senti- ment polarities of them as positive and negative sepa- rately. AESC is going to complete these two tasks from the given sentence. employs the two-stage method which first accom- plishes the AE and is followed by another model to perform SC, thus achieving the AESC (Yu et al., 2018; Hu et al., 2019; Fan et al., 2019). The second tries to fulfill the three subtasks by a more unified methodology which extracts the terms and their cor- responding polarities in a joint or interactive way (Liu et al., 2016; Wang et al., 2017; Li and Lam, 2017; Fan et al., 2018; He et al., 2019; Luo et al., 2019; Li et al., 2019; Peng et al., 2020; Chen and Qian, 2020; Chen et al., 2020; Wan et al., 2020). However, the above mentioned schemes perform ABSA task by stacking recurrent neural networks …

Content truncated to 500 words