This portal has been archived. Explore the next generation of this technology.
Discovering the Compositional Structure of Vector Representations with Role Learning Networks
lib:fcd2e068543c5a9e (v1.0.0)
Authors: Paul Soulos,Tom McCoy,Tal Linzen,Paul Smolensky
Where published: ICLR 2020 1
ArXiv: 1910.09113
Document: PDF DOI
Artifact development version: GitHub
Abstract URL: https://arxiv.org/abs/1910.09113v2
Neural networks are able to perform tasks that rely on compositional structure even though they lack obvious mechanisms for representing this structure. To analyze the internal representations that enable such success, we propose ROLE, a technique that detects whether these representations implicitly encode symbolic structure. ROLE learns to approximate the representations of a target encoder E by learning a symbolic constituent structure and an embedding of that structure into E's representational vector space. The constituents of the approximating symbol structure are defined by structural positions - roles - that can be filled by symbols. We show that when E is constructed to explicitly embed a particular type of structure (string or tree), ROLE successfully extracts the ground-truth roles defining that structure. We then analyze a GRU seq2seq network trained to perform a more complex compositional task (SCAN), where there is no ground truth role scheme available. For this model, ROLE successfully discovers an interpretable symbolic structure that the model implicitly uses to perform the SCAN task, providing a comprehensive account of the representations that drive the behavior of a frequently-used but hard-to-interpret type of model. 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 resulting output is changed in the way predicted by our analysis. Finally, we use ROLE to explore whether popular sentence embedding models are capturing compositional structure and find evidence that they are not; we conclude by suggesting how insights from ROLE can be used to impart new inductive biases to improve the compositional abilities of such models.
Where published: ICLR 2020 1
ArXiv: 1910.09113
Document: PDF DOI
Artifact development version: GitHub
Abstract URL: https://arxiv.org/abs/1910.09113v2
Neural networks are able to perform tasks that rely on compositional structure even though they lack obvious mechanisms for representing this structure. To analyze the internal representations that enable such success, we propose ROLE, a technique that detects whether these representations implicitly encode symbolic structure. ROLE learns to approximate the representations of a target encoder E by learning a symbolic constituent structure and an embedding of that structure into E's representational vector space. The constituents of the approximating symbol structure are defined by structural positions - roles - that can be filled by symbols. We show that when E is constructed to explicitly embed a particular type of structure (string or tree), ROLE successfully extracts the ground-truth roles defining that structure. We then analyze a GRU seq2seq network trained to perform a more complex compositional task (SCAN), where there is no ground truth role scheme available. For this model, ROLE successfully discovers an interpretable symbolic structure that the model implicitly uses to perform the SCAN task, providing a comprehensive account of the representations that drive the behavior of a frequently-used but hard-to-interpret type of model. 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 resulting output is changed in the way predicted by our analysis. Finally, we use ROLE to explore whether popular sentence embedding models are capturing compositional structure and find evidence that they are not; we conclude by suggesting how insights from ROLE can be used to impart new inductive biases to improve the compositional abilities of such models.
Relevant initiatives
Related knowledge about this paper
Reproduced results (crowd-benchmarking and competitions)
Artifact and reproducibility checklists
- ACM/cTuning artifact appendix and reproducibility checklist
- NeurIPS reproducibility checklist
- SIGPLAN's checklist for empirical evaluation
- Collective Knowledge (organizing research projects based on FAIR principles)