Meta-Learning for Domain Generalization in Semantic Parsing

Abstract

The importance of building semantic parsers which can be applied to new domains and generate programs unseen at training has long been acknowledged, and datasets testing outof-domain performance are becoming increasingly available. However, little or no attention has been devoted to learning algorithms or objectives which promote domain generalization, with virtually all existing approaches relying on standard supervised learning. In this work, we use a meta-learning framework which targets zero-shot domain generalization for semantic parsing. We apply a modelagnostic training algorithm that simulates zeroshot parsing by constructing virtual train and test sets from disjoint domains. The learning objective capitalizes on the intuition that gradient steps that improve source-domain performance should also improve target-domain performance, thus encouraging a parser to generalize to unseen target domains. Experimental results on the (English) Spider and Chinese Spider datasets show that the meta-learning objective significantly boosts the performance of a baseline parser.

Semantic-Discriminative Mixup for Generalizable Sensor-based Cross-domain Activity Recognition

“…Representation learning also includes two kinds of representative techniques: domain-invariant representation learning [13,25] which performs kernel, adversarial training, explicitly feature alignment between domains, or invariant risk minimization to learn domain-invariant representations and feature disentanglement [12,41] which tries to disentangle the features into domain-shared or domain-specific parts for better generalization. Learning strategy mainly has three kinds of methods: ensemble learning [16] which relies on the power of ensemble to learn a unified and generalized predictive function, meta-learning [2,30] which is based on the learning-to-learn mechanism to learn general knowledge by constructing meta-learning tasks to simulate domain shift, and gradient operation [11,28] which tries to learn generalized representations by directly operating on gradients. For more details, please refer to [33].…”

Domain Generalisation of NMT: Fusing Adapters with Leave-One-Domain-Out Training

“…Domain generalisation has been mostly studied in computer vision (Wang et al, 2021b). The main approaches include invariant feature learning (Li et al, 2018;Wang et al, 2020b), data augmentation (Wang et al, 2020a), and meta learning (Balaji et al, 2018;Wang et al, 2021a). Although domain mismatch is a known challenge in NMT (Müller et al, 2020), domain generalisation has just recently drawn attention with the introduction of zeroshot evaluation in WMT2020 Robustness shared task (Specia et al, 2020), but is still under-explored.…”

Code Generation for Unknown Libraries via Reading API Documentations

“…There are semantic parsing studies other than code generation that deal with out-of-domain settings. In the Text2SQL task, which generates SQL statements given a natural language intent and table schema as inputs, researchers have developed models that can generalize to table schemas that did not appear during training (Yu et al, 2018;Zhong et al, 2020;Suhr et al, 2020;Wang et al, 2021). Pasupat and Liang (2015) proposed a semantic parsing model for question answering on unknown tables.…”