Improving Aspect Sentiment Quad Prediction via Template-Order Data Augmentation

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7889 - 7900 December 7-11, 2022 ©2022 Association for Computational Linguistics Improving Aspect Sentiment Quad Prediction via Template-Order Data Augmentation Mengting Hu1 Yike Wu2 Hang Gao3∗ Yinhao Bai1 Shiwan Zhao† 1 College of Software, Nankai University 2 School of Journalism and Communication, Nankai University 3 Institute for Public Safety Research, Tsinghua University {mthu, wuyike}@nankai.edu.cn, gaohang@mail.tsinghua.edu.cn yinhao@mail.nankai.edu.cn, zhaosw@gmail.com Abstract Recently, aspect sentiment quad prediction (ASQP) has become a popular task in the field of aspect-level sentiment analysis. Previous work utilizes a predefined template to para- phrase the original sentence into a structure target sequence, which can be easily decoded as quadruplets of the form (aspect category, aspect term, opinion term, sentiment polarity). The template involves the four elements in a fixed order. However, we observe that this so- lution contradicts with the order-free property of the ASQP task, since there is no need to fix the template order as long as the quadruplet is extracted correctly. Inspired by the observa- tion, we study the effects of template orders and find that some orders help the generative model achieve better performance. It is hy- pothesized that different orders provide various views of the quadruplet. Therefore, we propose a simple but effective method to identify the most proper orders, and further combine mul- tiple proper templates as data augmentation to improve the ASQP task. Specifically, we use the pre-trained language model to select the or- ders with minimal entropy. By fine-tuning the pre-trained language model with these template orders, our approach improves the performance of quad prediction, and outperforms state-of- the-art methods significantly in low-resource settings1. 1 Introduction The aspect sentiment quad prediction (ASQP) task, aiming to extract aspect quadruplets from a review sentence, becomes popular recently (Zhang et al., 2021a; Cai et al., 2021). The quadruplet consists of four sentiment elements: 1) aspect category (ac) in- dicating the aspect class; 2) aspect term (at) which is the specific aspect description; 3) opinion term ∗Corresponding author. †Independent researcher. 1Experimental codes and data are available at: https: //github.com/hmt2014/AspectQuad. The restaurant is clean. ambience general is great because restaurant is clean Original sentence Target sequence ( , , , ac at ot sp) (ambience general, restaurant, clean, positive) ( , , , ) xac xat xot xsp (ambience general, restaurant, clean, great) Quadruplet Semantic quadruplet Fixed-order template is because is xac xsp xat xot Multiple target sequences Free-order template Oi([𝙰𝙲] xac, [𝙰𝚃] xat, [𝙾𝚃] xot, [𝚂𝙿] xsp); i ∈ [1,24] [𝙰𝙲] xac [𝙰𝚃] xat [𝙾𝚃] xot [𝚂𝙿] xsp [𝙰𝚃] xat [𝙰𝙲] xac [𝙾𝚃] xot [𝚂𝙿] xsp … Figure 1: An example sentence is paraphrased into a tar- get sequence with a fixed-order template (Zhang et al., 2021a). Our approach employs special markers to form free-order templates and generates multiple target se- quences. Oi is the i-th order permutation of the four elements. (ot) which is the opinion expression towards the as- pect; 4) sentiment polarity (sp) denoting the senti- ment class of …

Content truncated to 500 words