On the naturalness of software

By Premkumar T. Devanbu. 2012

1,021 citations recorded by the source corpus.

Subfields: Information Systems, Software, Code Generation & Program Synthesis

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Information Systems

In Software

In Code Generation & Program Synthesis

Citations in the corpus

Competition-level code generation with AlphaCode (2022, background): “Hindle et al. (2012) provided empirical evidence that code can be modeled by statistical 𝑛-gram language models, and capitalised on this to develop a simple code completion engine for Java.”
Program Synthesis with Large Language Models (2021, methodology): “Early work applied statistical n-gram models [Hindle et al., 2012, Allamanis and Sutton, 2013a] and neural networks [Maddison and Tarlow, 2014, Raychev et al., 2014] to code.”
Evaluating Large Language Models Trained on Code (2021, background): “Hindle et al. (2012) investigated n-gram language models of code, ﬁnding code to be more predictable than natural language.”
A Survey of Machine Learning for Big Code and Naturalness (2017, methodology): “[81] were the first to use a token-level n-gram LM for this purpose, using the previous n − 1 tokens to represent the completion context at each location.”
StarCoder: may the source be with you! (2023, methodology): “Language Models for Code Language models were initially applied to code by Hindle et al. (2012), but relied on n-gram models trained at comparatively small scale.”
A systematic evaluation of large language models of code (2022, background): “Recently, LMs have shown impressive performance in modeling also source code, written in programming languages (Hindle et al., 2016; Hellendoorn & Devanbu, 2017; Alon et al., 2020; Karampatsis et al., 2020).”
CodeXGLUE: A Machine Learning Benchmark Dataset for Code Understanding and Generation (2021, background): “Similarly, developers who are confused about what to write next can use code completion systems [57, 1, 56, 66, 67, 8, 27, 7] to automatically complete the following tokens based on the edits made to the code.”
code2seq: Generating Sequences from Structured Representations of Code (2018, background): “Several works modeled code as a sequence of tokens (Hindle et al., 2012; Iyer et al., 2016; Allamanis et al., 2016; Loyola et al., 2017), characters (Bielik et al.”
Structured Generative Models of Natural Source Code (2014, background): “We previously mentioned Hindle et al. (2012) and Allamanis and Sutton (2013), which tackle the same task as us but with simple NLP models.”
Learning to Represent Programs with Graphs (2017, methodology): “ture open-source software projects (cf. subsection 5.3). 2 RELATED WORK Our work builds upon the recent ﬁeld of using machine learning for source code artifacts (Allamanis et al., 2017). For example, Hindle et al. (2012); Bhoopchandet al. (2016) model the code as a sequence of tokens, while Maddison”
Learning to Generate Pseudo-Code from Source Code Using Statistical Machine Translation (T) (2015, neutral): “…this study: MeCab to tokenize Japanese sentences [29], pialign to train word alignments [22], KenLM to train Kneser-Ney smoothed language model [30], Moses to train and generate target sentences using the PBMT model [31], and Travatar to train and generate target sentences using T2SMT models [32].”
A Convolutional Attention Network for Extreme Summarization of Source Code (2016, methodology): “Rush et al. (2015) use an attention-based encoder to summarize sentences, but do not use convolution for their attention mechanism.”
IntelliCode compose: code generation using transformer (2020, methodology): “ation to detect programming language. Given N lang one-dimensional monolingual arrays of subtokens, the training dataset for the double-headed GPT-C is extracted according to an algorithm detailed in [7]. As seen, language speciﬁc control codes provide necessary supervision to constrain language gen”
code2vec: learning distributed representations of code (2018, methodology): “…Context-Free Grammars [Allamanis and Sutton 2014; Bielik et al. 2016; Gvero and Kuncak 2015; Maddison and Tarlow 2014], n-grams [Allamanis et al. 2014; Allamanis and Sutton 2013; Hindle et al. 2012; Nguyen et al. 2013; Raychev et al. 2014] have been used for programming languages in the past.”
DeepBugs: a learning approach to name-based bug detection (2018, background): “All these approaches exploit the availability of a large number of examples to learn from, e.g., in the form of publicly available code repositories, and that source code has regularities even across projects written by different developers [Hindle et al. 2012].”
Big Code != Big Vocabulary: Open-Vocabulary Models for Source Code (2020, methodology): “A global LM, trained in a cross-project setting, will perform better if it is adapted to a new project [44, 81].”