Sentiment Interpretable Logic Tensor Network for Aspect-Term Sentiment Analysis

Proceedings of the 29th International Conference on Computational Linguistics, pages 6705–6714 October 12–17, 2022. 6705 Sentiment Interpretable Logic Tensor Network for Aspect-Term Sentiment Analysis Bowen Zhang1, Xu Huang2, Zhichao Huang3, Hu Huang4, Baoquan Zhang2, Xianghua Fu1∗ and Liwen Jing1,5 ∗ 1College of Big Data and Internet, Shenzhen Technology University, Shenzhen, China 2 School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen, China 3 JD Intelligent Cities Research, Beijing, China 4 School of Cyber Science and Technology, University of Science and Technology of China, Hefei, China 5 Shenzhen innovation Institute (Xinstitute), Shenzhen, China Abstract Aspect-term sentiment analysis (ATSA) is an important task that aims to infer the sentiment towards the given aspect-terms. It is often re- quired in the industry that ATSA should be performed with interpretability, computational efficiency and high accuracy. However, such an ATSA method has not yet been developed. This study aims to develop an ATSA method that ful- fills all these requirements. To achieve the goal, we propose a novel Sentiment Interpretable Logic Tensor Network (SILTN). SILTN is inter- pretable because it is a neurosymbolic formal- ism and a computational model that supports learning and reasoning about data with a differ- entiable first-order logic language (FOL). To realize SILTN with high inferring accuracy, we propose a novel learning strategy called the two- stage syntax knowledge distillation (TSynKD). Using widely used datasets, we experimentally demonstrate that the proposed TSynKD is ef- fective for improving the accuracy of SILTN, and the SILTN has both high interpretability and computational efficiency. 1 Introduction 1.1 Motivation Aspect-term sentiment analysis (ATSA) is a fine- grained task in sentiment analysis, which aims is to recognize the sentiment polarity of the given aspect-term in a sentence (Zafra et al., 2019). Early research for ATSA was developed based on man- ually extracted features. For example, Poria et al. (2014) designed hand-crafted dependency rules to obtain aspect-related words, which are then fed into the machine learning methods to infer the sentiment polarity. With the development of deep learning, deep neural networks (DNNs) have dominated the study. Generally, the performance of DNNs is supe- rior to traditional machine learning methods when ∗ Corresponding authors: xianghuafu@sztu.edu.cn, ljing@connect.ust.hk the labeled training data is sufficient. More re- cently, there are two classes of methods that have received attention: (i) syntax-aware neural network (SaNN), which integrates syntax knowledge into attention-based neural network that increases the predictor’s performance and interpretability (Zhang et al., 2019; Wang et al., 2020; Li et al., 2020; Nguyen and Shirai, 2015); (ii) large pre-trained lan- guage model (PLM) for ATSA (e.g., Bert (Devlin et al., 2019; Song et al., 2019; Zeng et al., 2019; Dai et al., 2021)), which learns knowledge from large- scale corpus and stably exceeds the other baseline by a significant margin. Despite the effectiveness of prior work, ATSA in real-world remains a challenge for several rea- sons. First, DNNs usually perform as a “black box”, because they cannot explicitly explain the process of the analysis; therefore, they cannot be applied in cases …

Content truncated to 500 words