Discovering the Compositional Structure of Vector Representations with Role Learning Networks

Paul Soulos, Tom McCoy, Tal Linzen, Paul Smolensky
10/21/2019
cs.LGcs.CLstat.ML

Official code: psoulos/role-decomposition (11 stars, last commit 2020-02-11). 1 more implementation below.

Abstract

How can neural networks perform so well on compositional tasks even though they lack explicit compositional representations? We use a novel analysis technique called ROLE to show that recurrent neural networks perform well on such tasks by converging to solutions which implicitly represent symbolic structure. This method uncovers a symbolic structure which, when properly embedded in vector space, closely approximates the encodings of a standard seq2seq network trained to perform the compositional SCAN task. We verify the causal importance of the discovered symbolic structure by showing that, when we systematically manipulate hidden embeddings based on this symbolic structure, the model's output is changed in the way predicted by our analysis.

View on arXivView PDF

Code Implementations(2)

psoulos/role-decompositionOfficial100%
112PythonOct 20, 20196 years ago
iclr2020-anonymous1/role-learnerOfficial100%
41PythonSep 22, 20197 years ago

Similar Papers

Do Deep Neural Networks Capture Compositionality in Arithmetic Reasoning?

86% match2023

Break It Down: Evidence for Structural Compositionality in Neural Networks

86% match1 repo92023

Scaling can lead to compositional generalization

83% match2025

Functorial Neural Architectures from Higher Inductive Types

83% match2026

Compositional Generalization with Tree Stack Memory Units

83% match1 repo172019

Neuro-symbolic Syntactic Parsing: Shaping a Neural Network with the CYK Algorithm

82% match2026

Cite this paper

@article{soulos2019discovering,
  title  = {Discovering the Compositional Structure of Vector Representations with Role Learning Networks},
  author = {Paul Soulos and Tom McCoy and Tal Linzen and Paul Smolensky},
  year   = {2019},
  eprint = {1910.09113},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/1910.09113}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/1910.09113.svg)](https://codeofpaper.com/paper/1910.09113)

Discussion