MGIMN: Multi-Grained Interactive Matching Network for Few-shot Text Classification

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1937 - 1946 July 10-15, 2022 ©2022 Association for Computational Linguistics MGIMN: Multi-Grained Interactive Matching Network for Few-shot Text Classification Jianhai Zhang1, Mieradilijiang Maimaiti1, Gao Xing1, Yuanhang Zheng2,3,4, and Ji Zhang1,∗ 1Alibaba DAMO Academy 2Department of Computer Science and Technology, Tsinghua University, Beijing, China 3Institute for Artificial Intelligence, Tsinghua University, Beijing, China 4Beijing National Research Center for Information Science and Technology {tanfan.zjh,mieradilijiang.mea,gaoxing.gx,zj122146}@alibaba-inc.com zheng-yh19@mails.tsinghua.edu.cn Abstract Text classification struggles to generalize to unseen classes with very few labeled text in- stances per class. In such a few-shot learning (FSL) setting, metric-based meta-learning ap- proaches have shown promising results. Pre- vious studies mainly aim to derive a prototype representation for each class. However, they neglect that it is challenging-yet-unnecessary to construct a compact representation which expresses the entire meaning for each class. They also ignore the importance to capture the inter-dependency between query and the sup- port set for few-shot text classification. To deal with these issues, we propose a meta- learning based method MGIMN which per- forms instance-wise comparison followed by aggregation to generate class-wise matching vectors instead of prototype learning. The key of instance-wise comparison is the interac- tive matching within the class-specific context and episode-specific context. Extensive exper- iments demonstrate that the proposed method significantly outperforms the existing SOTA approaches, under both the standard FSL and generalized FSL settings. 1 Introduction Few-shot text classification has attracted consid- erable attention because of significant academic research value and practical application value (Gao et al., 2019; Yin, 2020; Brown et al., 2020; Bragg et al., 2021; Liu et al., 2021). Many efforts are de- voted towards different goals like generalization to new classes (Gao et al., 2019; Ye and Ling, 2019; Nguyen et al., 2020; Bragg et al., 2021), adapta- tion to new domains and tasks (Bansal et al., 2020; Brown et al., 2020; Schick and Schütze, 2020; Gao et al., 2020; Bragg et al., 2021). Low-cost gen- eralization to new classes is critical to deal with the growing long-tailed categories, which is com- mon for intention classification (Geng et al., 2019; ∗ Corresponding author: Ji Zhang Q 𝑆𝑛𝑘 𝑆𝑛 𝑆 Instance level Class level Episode level A A A Instance-aware Class-aware Episode-aware Contextual-aware Q 𝑆𝑛 𝑆 A A A Query-aware Class-aware Episode-aware Instance matching vector 𝑆𝑛𝑘 M Contextual-aware Query Figure 1: The example of multi-grained few-shot text classification. The “Snk, Sn” and “S” represent the in- stance level, class level and episode level interaction re- spectively. While “A” in the purple circle denotes align- ment between query and instances. The “’M” stands for matching operation. Krone et al., 2020) and relation classification (Han et al., 2018; Gao et al., 2019). To prevent over-fitting to few-shot new classes and avoid retraining the model when the class space changes, metric-based meta learning has become the major framework with significant results (Yin, 2020; Bansal et al., 2020). The core idea is that episode sampling is …

Content truncated to 500 words