TART: Improved Few-shot Text Classification Using Task-Adaptive Reference Transformation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 11014–11026 July 9-14, 2023 ©2023 Association for Computational Linguistics TART: Improved Few-shot Text Classification Using Task-Adaptive Reference Transformation Shuo Lei†, Xuchao Zhang‡, Jianfeng He†, Fanglan Chen†, Chang-Tien Lu† †Department of Computer Science, Virginia Tech, Falls Church, VA, USA ‡Microsoft, Redmond, WA, USA {slei,jianfenghe,fanglanc,ctlu}@vt.edu xuchaozhang@microsoft.com Abstract Meta-learning has emerged as a trending tech- nique to tackle few-shot text classification and achieve state-of-the-art performance. How- ever, the performance of existing approaches heavily depends on the inter-class variance of the support set. As a result, it can perform well on tasks when the semantics of sampled classes are distinct while failing to differentiate classes with similar semantics. In this paper, we propose a novel Task-Adaptive Reference Transformation (TART) network, aiming to enhance the generalization by transforming the class prototypes to per-class fixed refer- ence points in task-adaptive metric spaces. To further maximize divergence between trans- formed prototypes in task-adaptive metric spaces, TART introduces a discriminative ref- erence regularization among transformed pro- totypes. Extensive experiments are conducted on four benchmark datasets and our method demonstrates clear superiority over the state- of-the-art models in all the datasets. In partic- ular, our model surpasses the state-of-the-art method by 7.4% and 5.4% in 1-shot and 5-shot classification on the 20 Newsgroups dataset, respectively. Our code is available at https: //github.com/slei109/TART 1 Introduction Deep learning has achieved great success in many fields but a deficiency of supervised data is often experienced in real-world NLP applications. Few- shot text classification aims to perform classifica- tion with a limited number of training instances, which is crucial for many applications but remains to be a challenging task. Existing approaches for few-shot text classifi- cation mainly fall into two categories: i) prompt- based learning (Brown et al., 2020; Gao et al., 2021; Wang et al., 2021), which utilizes Pre-trained Lan- guage Models (PLMs) to generate a textual answer in response to a given prompt. Although produc- ing promising results, these methods suffer from Class Testing Sample Task 1: Support class: 1,2,3,4 Task 2: Support class: 1,2,3,5 MLADA Ours MLADA Ours 1 Animal photos of the week: baby tiger goes for a swim. 1 1 1 1 2 Twitter helps confirm X-shaped bulge at Center of Milky Way. 4 2 2 2 3 Toronto van attack suspect's Facebook post praised misogynist mass killer. 4 3 2 3 4 Apple just solved one of the iphone's most harmful features. 2 4 - - 5 Apple fritter season is here, and so are the recipes you’ll need. - - 5 5 Class 1: Environment Class 2: Science Class 3: World News Class 4: Tech Class 5: Taste Figure 1: Prediction results of example tasks with different inter-class variance on the Huffpost dataset. MLADA (Han et al., 2021) performs well on the task with high inter-class variance (e.g., Task 2: Environment, Science, World News, Taste), while it fails to distinguish the samples from a task with low inter-class variance …

Content truncated to 500 words