Generative Data Augmentation for Aspect Sentiment Quad Prediction

Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 128–140 July 13-14, 2023 ©2023 Association for Computational Linguistics Generative Data Augmentation for Aspect Sentiment Quad Prediction An Wang† Junfeng Jiang‡ Youmi Ma† Ao Liu† Naoaki Okazaki† †Tokyo Institute of Technology ‡The University of Tokyo an.wang@nlp.c.titech.ac.jp jiangjf@is.s.u-tokyo.ac.jp {youmi.ma@nlp., ao.liu@nlp., okazaki@}c.titech.ac.jp Abstract Aspect sentiment quad prediction (ASQP) an- alyzes the aspect terms, opinion terms, senti- ment polarity, and aspect categories in a text. One challenge in this task is the scarcity of data owing to the high annotation cost. Data augmentation techniques are commonly used to address this issue. However, existing ap- proaches simply rewrite texts in the training data, restricting the semantic diversity of the generated data and impairing the quality due to the inconsistency between text and quads. To address these limitations, we augment quads and train a quads-to-text model to generate cor- responding texts. Furthermore, we designed novel strategies to filter out low-quality data and balance the sample difficulty distribution of the augmented dataset. Empirical studies on two ASQP datasets demonstrate that our method outperforms other data augmentation methods and achieves state-of-the-art perfor- mance on the benchmarks.1 1 Introduction Aspect-based sentiment analysis (ABSA) aims to mine opinions expressed regarding specific aspects of a given text. Recently, Zhang et al. (2021a) pro- posed a challenging compound subtask of ABSA called aspect sentiment quad prediction (ASQP), which predicts four kinds of elements (aspect cate- gory, aspect term, opinion term, sentiment polarity) as quadruplets (quads). A single text may contain multiple quads. For example, the text “The pizza is delicious but expensive.” mentions one aspect term (pizza) and two opinion terms (delicious and expensive). Because these two opinions are related to the same aspect, the text includes two quads: (taste, pizza, delicious, positive) and (price, pizza, expensive, negative). Traditional methods (Cai et al., 2020; Wan et al., 2020; Cai et al., 2021) address such compound sub- 1The source code is available at https://github. com/AnWang-AI/AugABSA It is one the nicest outdoor restaurants I have ever seen in NY. It is one of the most beautiful outdoor restaurants I have ever seen in NY. It is one the nicest restaurants I have ever seen in NY. Random Deletion: Back Translation: It is one the best outdoor restaurants I have ever seen in NY. Synonym Replacement: [ambience general, outdoor restaurants, nicest, positive] Source Text: Label: Augmented Text (Previous Methods): Great space for meetings, but the outdoor restaurants is nothing special. Our Method: Augmented Label: [ambience general, space, Great, positive], [ambience general, outdoor restaurants, nothing special, positive] Augmented Text: Figure 1: Examples of text data augmentation meth- ods. We observe that the augmented texts from previous methods fail to include all spans in the label and the augmented texts are semantically very similar to the source text. Our method addresses these problems by generating texts from augmented labels. tasks of ABSA in a discriminative manner. Recent studies (Zhang et al., 2021b; Hu et al., 2022) have primarily concentrated on sequence-to-sequence …

Content truncated to 500 words