Assign continuous vectors to logical and algebraic symbolic expressions in such a way that semantically equivalent, but syntactically diverse expressions are assigned to identical (or highly similar) continuous vectors.
ICML talk slides and abstract can be found here.
The data used in the paper can be downloaded from here.
We provide the source code on GitHub.
In the following links you can find some t-SNE visualizations of the SemVecs learned by the neural equivalence network. Due to their size constraints (out-of-memory error when computing pairwise distances), some datasets are excluded or partially included. You can zoom and drag the visualization. Hovering over an element, you can see the expression and the equivalence class it belongs in. The lines appearing when hovering over an expression connect all points that belong to the same equivalence class.
Please cite as
@article{allamanis2016learning,
title={Learning Continuous Semantic Representations of Symbolic Expressions},
author={Allamanis, Miltiadis and Chanthirasegaran, Pankajan and Kohli, Pushmeet and Sutton, Charles},
journal={arXiv preprint arXiv:1611.01423},
year={2016}
}