A general-purpose neural semantic parser for mapping natural language queries into machine executable code
Switch branches/tags
Nothing to show
Clone or download
Latest commit 925ea64 Oct 12, 2018

README.md

TranX

A general-purpose Transition-based abstract syntaX parser that maps natural language queries into machine executable source code (e.g., Python) or logical forms (e.g., lambda calculus).

System Architecture

Details could be found in this technical report. To cope with different domain specific logical formalisms (e.g., SQL, Python, lambda-calculus, prolog, etc.), TranX uses abstract syntax trees (ASTs) defined in the Abstract Syntax Description Language (ASDL) as intermediate meaning representation.

Sysmte Architecture

Figure 1 gives a brief overview of the system.

  1. TranX first employs a transition system to transform a natural language utterance into a sequence of tree-constructing actions, following the input grammar specification of the target formal language. The grammar specification is provided by users in textual format (e.g., asdl/lang/py_asdl.txt for Python grammar).

  2. The tree-constructing actions produce an intermediate abstract syntax tree. TranX uses ASTs defined under the ASDL formalism as general-purpose, intermediate meaning representations.

  3. The intermediate AST is finally transformed to a domain-specific representation (e.g., Python source code) using customly-defined conversion functions.

File Structure tranX is mainly composed of two components:

  1. A general-purpose transition system that defines the generation process of an AST z using a sequence of tree-constructing actions a_0, a_1, ..., a_T.
  2. A neural network that computes the probability distribution over action sequences, conditional on the natural language query x, p(a_0, a_1, ..., a_T | x).

These two components are implemented in the following two folders, respectively:

  • asdl defines a general-purpose transition system based on the ASDL formalism, and its instantiations in different programming languages and datasets. The transition system defines how an AST is constructed using a sequence of actions. This package can be used as a standalone library independent of tranX. See Section 2.2 of the technical report for details.
  • model contains the neural network implementation of the transition system defined in asdl, which computes action probabilities using neural networks. See Section 2.3 of the technical report for details.

Additionally, the components folder contains helper functions and classes (e.g., vocabulary).

Supported Language

TranX officially supports the following grammatical formalism and datasets. More languages (C#) are coming!

Language ASDL Specification Example Datasets
Python asdl/lang/py_asdl.txt Django (Oda et al., 2015)
lambda calculus asdl/lang/lambda_asdl.txt ATIS, GeoQuery (Zettlemoyer and Collins, 2005)
prolog asdl/lang/prolog_asdl.txt Jobs (Zettlemoyer and Collins, 2005)
SQL asdl/lang/sql/sql_asdl.txt WikiSQL (Zhong et al., 2017)

Usage

Conda Environments

TranX supports both Python 2.7 and 3.5. Please note that some datasets only support Python 2.7 (e.g., Django) or Python 3+ (e.g., WikiSQL). We provide example conda environments (data/env/(py2torch3cuda9.yml|py3torch3cuda9.yml)) for both Python versions. You can export the enviroments using the following command:

conda env create -f data/env/(py2torch3cuda9.yml|py3torch3cuda9.yml)

Note The conda enviroments are generated on a Ubuntu 16.04 machine. If you are unable to import the enviroment with error message like cannot found package numpy=1.14.3=py36h14a74c5_0, please try removing the sha after the version number (e.g., =py36h14a74c5_0), since it might be different on different platforms. We keep the detailed version number with sha to ensure reproducibility.

Example Training Script

git clone https://github.com/pcyin/tranX
cd tranX

. pull_data.sh  # get datasets and training scripts for supported languages

The scripts folder contains scripts to train TranX on example datasets. For example, to train on the Django dataset, simply run:

. scripts/django/train.sh 0  # start training on Django dataset with seed 0

Using the provided conda environment, it achieves 73.9% test accuracy on a Ubuntu 16.04 machine with GTX1080 GPU.

FAQs

How to adapt to a new programming language or logical form?

You need to implement the TransitionSystem class with a bunch of custom functions which (1) convert between domain-specific logical forms and intermediate ASTs used by TranX, (2) predictors which check if a hypothesis parse if correct during beam search decoding. You may take a look at the examples in asdl/lang/*.

How to generate those pickled datasets (.bin files)?

Please refer to asdl/lang/*/dataset.py for code snippets that converts a dataset into pickled files.

Reference

TranX is described/used in the following two papers:

@inproceedings{yin18emnlpdemo,
    title = {{TRANX}: A Transition-based Neural Abstract Syntax Parser for Semantic Parsing and Code Generation},
    author = {Pengcheng Yin and Graham Neubig},
    booktitle = {Conference on Empirical Methods in Natural Language Processing (EMNLP) Demo Track},
    year = {2018}
}

@inproceedings{yin18acl,
    title = {Struct{VAE}: Tree-structured Latent Variable Models for Semi-supervised Semantic Parsing},
    author = {Pengcheng Yin and Chunting Zhou and Junxian He and Graham Neubig},
    booktitle = {The 56th Annual Meeting of the Association for Computational Linguistics (ACL)},
    url = {https://arxiv.org/abs/1806.07832v1},
    year = {2018}
}

We are also grateful to the following papers that inspire this work :P

Abstract Syntax Networks for Code Generation and Semantic Parsing.
Maxim Rabinovich, Mitchell Stern, Dan Klein.
in Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2017

The Zephyr Abstract Syntax Description Language.
Daniel C. Wang, Andrew W. Appel, Jeff L. Korn, and Christopher S. Serra.
in Proceedings of the Conference on Domain-Specific Languages, 1997