Relevance Transformer: Generating Concise Code Snippets with Relevance Feedback

Abstract

Tools capable of automatic code generation have the potential to augment programmer's capabilities. While straightforward code retrieval is incorporated into many IDEs, an emerging area is explicit code generation. Code generation is currently approached as a Machine Translation task, with Recurrent Neural Network (RNN) based encoder-decoder architectures trained on code-description pairs. In this work we introduce and study modern Transformer architectures for this task. We further propose a new model called the Relevance Transformer that incorporates external knowledge using pseudo-relevance feedback. The Relevance Transformer biases the decoding process to be similar to existing retrieved code while enforcing diversity. We perform experiments on multiple standard benchmark datasets for code generation including Django, Hearthstone, and CoNaLa. The results show improvements over state-ofthe-art methods based on BLEU evaluation. The Relevance Transformer model shows the potential of Transformer-based architectures for code generation and introduces a method of incorporating pseudo-relevance feedback during inference.

A Survey on Artificial Intelligence for Source Code: A Dialogue Systems Perspective

“…(Liguori et al, 2021a)). Both exact match and averaged token level BLEU scores have been extensively used in evaluating code generation models (Liguori et al, 2021a,b;Oda et al, 2015b;Ling et al, 2016;Gemmell et al, 2020). It is becoming increasingly important to note the drawback of using BLEU to evaluate code generation systems.…”

“…(Chen Jr and Bunescu, 2019;Jia and Liang, 2016;Yin and Neubig, 2017a;Rabinovich et al, 2017;Ling et al, 2016;Sun et al, 2019a)), and transformer-based architectures (e.g. (Kacupaj et al, 2021;Ferraro and Suominen, 2020;Shen et al, 2019;Sun et al, 2019b;Gemmell et al, 2020;Kusupati and Ailavarapu)). In the late 1970's, Hendrix et al (1978) pioneer the task of interfacing with databases using natural language.…”

“…Akinbou et al ( 2021) use a standard transformer architecture and achieve 8.97 BLEU on the CoNaLa dataset with a 42% syntax error rate. Gemmell et al (2020) utilize a transformer architecture to translate natural language intents to executable snippet representations in Python. The authors further improve upon the baseline transformer model by introducing pseudo-relevance feedback during the decoding process.…”

Can We Generate Shellcodes via Natural Language? An Empirical Study

“…Ling et al [56] and Yin and Neubig [88] proposed a novel neural architecture for code generation, while Xu et al [84] incorporated pre-training and fine-tuning of a model to generate Python snippets from natural language using the CoNaLa dataset [87]. Furthermore, Gemmell et al [30] used a transformer architecture with relevance feedback for code generation, and reported improvements over state-of-the-art on several datasets. There also exist approaches that perform the reverse task, i.e., generating natural language from code.…”

PGT: Pseudo Relevance Feedback Using a Graph-Based Transformer

“…Most Transformer-based rerankers are applied to individual query-document pairs. Some research explores jointly modeling multiple top retrieved documents in a Transformer architecture for question clarification [11], question answering [10,14] or code generation [8]. The effectiveness of using top retrieved documents in Transformer rerankers remains to be studied.…”