Coupling Retrieval and Meta-Learning for Context-Dependent Semantic Parsing

Abstract

In this paper, we present an approach to incorporate retrieved datapoints as supporting evidence for context-dependent semantic parsing, such as generating source code conditioned on the class environment. Our approach naturally combines a retrieval model and a meta-learner, where the former learns to find similar datapoints from the training data, and the latter considers retrieved datapoints as a pseudo task for fast adaptation. Specifically, our retriever is a context-aware encoderdecoder model with a latent variable which takes context environment into consideration, and our meta-learner learns to utilize retrieved datapoints in a model-agnostic meta-learning paradigm for fast adaptation. We conduct experiments on CONCODE and CSQA datasets, where the context refers to class environment in JAVA codes and conversational history, respectively. We use sequence-to-action model as the base semantic parser, which performs the state-of-the-art accuracy on both datasets. Results show that both the context-aware retriever and the meta-learning strategy improve accuracy, and our approach performs better than retrieve-and-edit baselines.

Adversarial Training for Code Retrieval with Question-Description Relevance Regularization

“…Our goal is to learn a QC model, denoted as f QC θ , that retrieves the highest score code snippets for an input question: arg max c ∈{c}∪{c − } f QC θ (q, c ). Note that at testing time, the trained QC model f QC can be used to retrieve code snippets from any code bases, unlike the group of QC methods (Hayati et al, 2018;Hashimoto et al, 2018;Guo et al, 2019) relying on the availability of NL descriptions of code.…”

“…Since there is only one question for each sample in the QC datasets, we apply a standard code summarization tool (Iyer et al, 2016) to generate code descriptions to match with input questions. • vMF-VAE (Guo et al, 2019). This is similar to EditDist, but a vMF Variational Autoencoder (Xu and Durrett, 2018) is separately trained to embed questions and code descriptions into latent vector distributions, whose distance is then measured by KL-divergence.…”

“…This methodology has been widely adopted as a component in the retrieve-and-edit code generation literature. For example, heuristic methods such as measuring edit distance (Hayati et al, 2018) or comparing code type and length are used, and separate question latent representations (Hayati et al, 2018;Guo et al, 2019) are learned. Our work shares with them the idea to exploit QD relevance, but we use QD relevance in a novel way to regularize the adversarial learning of QC models.…”

“…Moreover, even if the retrieved code snippets do not perfectly match the NL question, editing them is often much easier than generating a code snippet from scratch. For example, the retrieve-and-edit paradigm (Hayati et al, 2018;Hashimoto et al, 2018;Guo et al, 2019) for code generation has attracted growing attention recently, which first employs a code retriever to find the most relevant code snippets for a given question, and then edit them via a code generation model. Previous work has shown that code retrieval performance can significantly affect the final generated results in such scenarios.…”

“…There have been two groups of work on code retrieval: (1) One group of work (e.g., the recent retrieve-and-edit work (Hashimoto et al, 2018;Guo et al, 2019)) assumes each code snippet is associated with NL descriptions and retrieves code snippets by measuring the relevance between such descriptions and a given question. (2) The other group of work (e.g., CODENN (Iyer et al, 2016) and Deep Code Search (Gu et al, 2018)) directly measures the relevance between a question and a code snippet.…”