NL2Bash

Overview

This fork of the NL2Bash repository was used to recreate the results from the original paper on the DAS-5 cluster for Nedim Azar's master's thesis.

The train_models.sh script can be modified to run the models on DAS-5.
The evalueate.sh script can be used to calculate performance metrics on
The fine_tuned_model_outputs directory contains outputs of fine-tuned GPT-3 models for evaluation on this task. Babbage and Curie outperdorm the previously presented models.

This repository contains the data and source code release of the paper: NL2Bash: A Corpus and Semantic Parser for Natural Language Interface to the Linux Operating System.

Specifically, it contains the following components:

A set of ~10,000 bash one-liners collected from websites such as StackOverflow paired with their English descriptions written by Bash programmers.
Tensorflow implementations of the following translation models:
- the standard Seq2Seq and CopyNet models
- a stage-wise NL⟶Bash model using argument filling heuristics (Lin et. al. 2017).
A Bash command parser which parses a Bash command into an abstractive syntax tree, developed on top of bashlex.
A set of domain-specific natural language processing tools, including a regex-based sentence tokenizer and a domain specific named entity recognizer.

You may visit http://tellina.rocks to interact with our pretrained model.

🆕 Apr 24, 2020 The dataset data/bash is separately licensed under MIT license.

Data Statistics

Our corpus contains a diverse set of Bash utilities and flags: 102 unique utilities, 206 unique flags and 15 reserved tokens. (Browse the raw data collection here.)

In our experiments, the set of ~10,000 NL-bash command pairs are splitted into train, dev and test sets such that neither a natural language description nor a Bash command appears in more than one split.

The statistics of the data split is tabulated below. (A command template is defined as a Bash command with all of its arguments replaced by their semantic types.)

Split	Train	Dev	Test
# pairs	8,090	609	606
# unique NL	7,340	549	547
# unique command	6,400	599	XX
# unique command template	4,002	509	XX

The frequency of the top 50 most frequent Bash utilities in the corpus is illustrated in the following diagram.

Leaderboard

Manually Evaluated Translation Accuracy

Top-k full command accuracy and top-k command template accuracy judged by human experts. Please refer to section 4 of the paper for the exact procedures we took to run manual evaluation.

Model	F-Acc-Top1	F-Acc-Top3	T-Acc-Top1	T-Acc-Top3
Sub-token CopyNet (this work)	0.36	0.45	0.49	0.61
Tellina (Lin et. al. 2017)	0.27	0.32	0.53	0.62

⚠️ If you plan to run manual evaluation yourself, please refer to "Notes on Manual Evaluation" for issues you should pay attention to.

Automatic Evaluation Metrics

In addition, we also report character-based BLEU and a self-defined template matching score as the automatic evaluation metrics used to approximate the true translation accuracy. Please refer to appendix C of the paper for the metrics definitions.

Model	BLEU-Top1	BLEU-Top3	TM-Top1	TM-Top3
Sub-token CopyNet (this work)	50.9	58.2	0.574	0.634
Tellina (Lin et. al. 2017)	48.6	53.8	0.625	0.698

Run Experiments

Install TensorFlow

To reproduce our experiments, please install TensorFlow 2.0. The experiments can be reproduced on machines with or without GPUs. However, training with CPU only is extremely slow and we do not recommend it. Inference with CPU only is slow but tolerable.

We suggest following the official instructions to install the library. The code has been tested on Ubuntu 16.04 + CUDA 10.0 + CUDNN 7.6.3.

Environment Variables & Dependencies

Once TensorFlow is installed, use the following commands to set up the Python path and main experiment dependencies.

export PYTHONPATH=`pwd`

(sudo) make

Change Directory

Then enter the scripts directory.

cd scripts

Data filtering, split and pre-processing

Run the following command. This will clean the raw NL2Bash corpus and apply filtering, create the train/dev/test splits and preprocess the data into the formats taken by the Tensorflow models.

make data

To change the data-processing workflow, go to data and modify the utility scripts.

Train models

make train

Evaluate models

We provide evaluation scripts to evaluate the performance of any new model.

To do so please save your model output to a file (example). We assume the file is of the following format:

1. The i-th line of the file contains predictions for example i in the dataset.
2. Each line contains top-k predictions separated by "|||".

Then get the evaluation results using the following script

Manual

Dev set evaluation

./bash-run.sh --data bash --prediction_file <path_to_your_model_output_file> --manual_eval

Test set evaluation

./bash-run.sh --data bash --prediction_file <path_to_your_model_output_file> --manual_eval --test

Automatic

Dev set evaluation

./bash-run.sh --data bash --prediction_file <path_to_your_model_output_file> --eval

Test set evaluation

./bash-run.sh --data bash --prediction_file <path_to_your_model_output_file> --eval --test

Generate evaluation table using pre-trained models

Decode the pre-trained models and print the evaluation summary table.

make decode

Skip the decoding step and print the evaluation summary table from the predictions saved on disk.

make gen_manual_evaluation_table

By default, the decoding and evaluation steps will print sanity checking messages. You may set verbose to False in the following source files to suppress those messages.

encoder_decoder/decode_tools.py
eval/eval_tools.py

Notes on Manual Evaluation

In our experiment, we conduct manual evaluation as the correctness of a Bash translation cannot simply be determined by mapping it to a set of ground truth. We suggest the following practices for future work to generate comparable results and to accelerate the development cycle.

If you plan to run your own manual evaluation, please annotate the output of both your system(s) and the baseline systems you compared to. This is to ensure that the newly proposed system(s) and the baselines are judged by the same group of annotators.
If you run manual evaluation, please release the examples annotated with their annotations. This helps others to replicate the results and reuse these annotations.
During model development you could annotate a small subset of the dev examples (50-100 is likely enough) to estimate the true dev set accuracy. We released a script which saves any previous annotations and opens a commandline interface for judging any unseen predictions (manual_eval.md).

The motivation for the practices above is detailed in issue #6.

Citation

If you use the data or source code in your work, please cite

@inproceedings{LinWZE2018:NL2Bash, 
  author = {Xi Victoria Lin and Chenglong Wang and Luke Zettlemoyer and Michael D. Ernst}, 
  title = {NL2Bash: A Corpus and Semantic Parser for Natural Language Interface to the Linux Operating System}, 
  booktitle = {Proceedings of the Eleventh International Conference on Language Resources
               and Evaluation {LREC} 2018, Miyazaki (Japan), 7-12 May, 2018.},
  year = {2018} 
}

Changelog

Apr 24, 2020 The dataset data/bash is separately licensed under the terms of the MIT license.
Oct 20, 2019 release standard evaluation scripts
Oct 20, 2019 update to Tensorflow 2.0

nedimazar/nl2bash