Unsupervised Data Augmentation for Aspect Based Sentiment Analysis

Proceedings of the 29th International Conference on Computational Linguistics, pages 6746–6751 October 12–17, 2022. 6746 Unsupervised Data Augmentation for Aspect Based Sentiment Analysis David Z. Chen Adam Faulkner Sahil Badyal Capital One Servicing Intelligence, NLP {david.chen2, adam.faulkner, sahil.badyal} @capitalone.com Abstract Recent approaches to Aspect-based Senti- ment Analysis (ABSA) take a co-extraction approach to this span-level classification task, performing the subtasks of aspect term extraction (ATE) and aspect senti- ment classification (ASC) simultaneously. In this work, we build on recent progress in applying pre-training to this co-extraction task with the introduction of an adapta- tion of Unsupervised Data Augmentation (UDA) in semi-supervised learning. As originally implemented, UDA cannot ac- commodate span-level classification since it relies on advanced data augmentation techniques, such as backtranslation, that alter the sequence lengths of the original data and cause index mismatches. We intro- duce an adaptation of UDA using Masked Language Model (MLM) unmasking that accommodates this index-match constraint and test the approach on standard ABSA benchmark datasets. We show that sim- ple augmentations applied to modest-sized datasets along with consistency training lead to competitive performance with the current ABSA state-of-the-art in the restau- rant and laptop domains using only 75% of the training data. 1 Introduction Aspect-based Sentiment Analysis (ABSA) is a subset of Sentiment Analysis (SA), operating at the phrase- rather then sentence- or document- level. As with other forms of SA, the goal is to determine the sentiment associated with a given text segment, though, in the case of ABSA, these phrasal segments are typically “aspects” or features associated with products, services, or experiences, such as “waitstaff” or “ambience.” As with other span-level classification tasks, such as Named Entity Recognition (NER), a major challenge of ABSA is class imbalance, as the majority of token labels typically refer to non-aspect terms (Luo et al., 2020) and the terms themselves are of inconsistent phrase- level categories. This introduces considerable variance in aspect term labels and makes it difficult for models to effectively generalize to example terms outside those explicitly shown in the training data. 1.1 Related work In related span-level tasks, previous work has shown that a joint/collapsed approach to en- tity and sentiment co-extraction out-performs a pipelined approach (Mitchell et al., 2013; Zhang et al., 2015). A joint approach refers to assigning two sets of tags, term and polar- ity, to each example, and a collapsed approach collapses the term and polarity tags into one term-polarity tag for each token. While there are merits to both approaches, we adopt a collapsed approach as it requires a simpler clas- sifier architecture. Pre-training and pre-trained language mod- els (LMs) have been shown to provide state- of-the-art performance on many tasks within NLP. Applying these approaches to the ABSA task, Li et al. (Li et al., 2019b) and Luo et al. (Luo et al., 2020) have achieved state-of-the-art performance on restaurant and laptop reviews using pre-trained LM’s and LM’s with post- training (PTR), respectively. These results highlight the benefit of leveraging unlabeled data (with pre-training and post-training). …

Content truncated to 500 words