Learning to Represent Edits

Authors: Pengcheng Yin, Graham Neubig, Miltiadis Allamanis, Marc Brockschmidt, Alexander L. Gaunt

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our evaluation yields promising results that suggest that our neural network models learn to capture the structure and semantics of edits. We experiment on natural language and source code edit data.
Researcher Affiliation Collaboration Pengcheng Yin , Graham Neubig Language Technology Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {pcyin,gneubig}@cs.cmu.edu Miltiadis Allamanis, Marc Brockschmidt, Alexander L. Gaunt Microsoft Research Cambridge, CB1 2FB, United Kingdom {miallama,mabrocks,algaunt}@microsoft.com
Pseudocode No The paper describes the models and methods using text and diagrams, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes (c) we create a new source code edit dataset, and release the data extraction code at https://github.com/Microsoft/msrc-dpu-learning-to-represent-edits and the data at http://www.cs.cmu.edu/ pengchey/githubedits.zip.
Open Datasets Yes We use the Wiki Atomic Edits (Faruqui et al., 2018) dataset... To obtain a dataset for source code, we clone a set of 54 C# projects on Git Hub and collected a Git Hub Edits dataset... and release the data at http://www.cs.cmu.edu/ pengchey/githubedits.zip.
Dataset Splits Yes We sampled 1040K edits from the English insertion portion of the dataset and split the samples into 1000K/20K/20K train-valid-test sets. ... We split the dataset into 91,372 / 10,176 / 10,176 samples as train/valid/test sets.
Hardware Specification No The paper does not specify any particular hardware (e.g., GPU models, CPU types, or cloud instances) used for running the experiments.
Software Dependencies No The paper describes model architectures and hyperparameters (e.g., 'bidirectional LSTM', 'edit representation size of 512', 'hidden states of encoding LSTMs is 128'), but does not list specific software dependencies with version numbers (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes Throughout the evaluation we use a fixed size of 512 for edit representations. The size of word embeddings and hidden states of encoding LSTMs is 128. The dimensionality of the decoding LSTM is set to 256. Details of model configuration can be found in Sect. A. ... For the graph-based edit encoder, we used a two-layer graph neural network, with 5 information propagation steps at each layer. During training, we performed early stopping, and choose the best model based on perplexity scores on development set. During testing, we decode a target element x+ using a beam size of 5.