Bidirectional Generative Framework for Cross-domain Aspect-based Sentiment Analysis

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12272–12285 July 9-14, 2023 ©2023 Association for Computational Linguistics Bidirectional Generative Framework for Cross-domain Aspect-based Sentiment Analysis Yue Deng ∗ 1,2 Wenxuan Zhang †1 Sinno Jialin Pan2,3 Lidong Bing1 1DAMO Academy, Alibaba Group 2 Nanyang Technological University, Singapore 3Chinese University of Hong Kong {yue.deng, saike.zwx, l.bing}@alibaba-inc.com sinnopan@cuhk.edu.hk Abstract Cross-domain aspect-based sentiment analysis (ABSA) aims to perform various fine-grained sentiment analysis tasks on a target domain by transferring knowledge from a source domain. Since labeled data only exists in the source do- main, a model is expected to bridge the domain gap for tackling cross-domain ABSA. Though domain adaptation methods have proven to be effective, most of them are based on a discrim- inative model, which needs to be specifically designed for different ABSA tasks. To offer a more general solution, we propose a unified bidirectional generative framework to tackle various cross-domain ABSA tasks. Specifi- cally, our framework trains a generative model in both text-to-label and label-to-text directions. The former transforms each task into a uni- fied format to learn domain-agnostic features, and the latter generates natural sentences from noisy labels for data augmentation, with which a more accurate model can be trained. To in- vestigate the effectiveness and generality of our framework, we conduct extensive exper- iments on four cross-domain ABSA tasks and present new state-of-the-art results on all tasks. Our data and code are publicly available at https://github.com/DAMO-NLP-SG/BGCA. 1 Introduction Aspect-based sentiment analysis (ABSA) is the task of analyzing people’s sentiments at the aspect level. It often involves several sentiment elements, including aspects, opinions, and sentiments (Liu, 2012; Zhang et al., 2022). For instance, given the sentence "The apple is sweet.", the aspect is ap- ple, its opinion is sweet, and the corresponding sentiment polarity is Positive. ABSA has attracted increasing attention in the last decade, and various tasks have been proposed to extract either single or ∗ Yue Deng is under the Joint PhD Program between Alibaba and Nanyang Technological University. † Wenxuan Zhang is the corresponding author. multiple sentiment elements under different scenar- ios. For example, aspect sentiment classification (ASC) predicts the sentiment polarity of a given aspect target (Chen et al., 2017; Li et al., 2018a; Xu et al., 2020a) and aspect term extraction (ATE) extracts aspects given the sentence (Li et al., 2018b; Liu et al., 2015), while aspect sentiment triplet ex- traction (ASTE) predicts all three elements in the triplet format (Peng et al., 2020; Xu et al., 2021). The main research line of ABSA focuses on solving various tasks within a specific domain. However, in real-world applications, such as E- commerce websites, there often exist a wide vari- ety of domains. Existing methods often struggle when applying models trained in one domain to un- seen domains, due to the variability of aspect and opinion expressions across different domains (Ding et al., 2017; Wang and Pan, 2018, 2019). Moreover, manually labeling data for each domain can be …

Content truncated to 500 words