We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

cs

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: A causal framework for explaining the predictions of black-box sequence-to-sequence models

Abstract: We interpret the predictions of any black-box structured input-structured output model around a specific input-output pair. Our method returns an "explanation" consisting of groups of input-output tokens that are causally related. These dependencies are inferred by querying the black-box model with perturbed inputs, generating a graph over tokens from the responses, and solving a partitioning problem to select the most relevant components. We focus the general approach on sequence-to-sequence problems, adopting a variational autoencoder to yield meaningful input perturbations. We test our method across several NLP sequence generation tasks.
Comments: 12 Pages, EMNLP 2017
Subjects: Machine Learning (cs.LG)
Cite as: arXiv:1707.01943 [cs.LG]
  (or arXiv:1707.01943v3 [cs.LG] for this version)

Submission history

From: David Alvarez-Melis [view email]
[v1] Thu, 6 Jul 2017 19:36:43 GMT (394kb,D)
[v2] Tue, 25 Jul 2017 16:31:23 GMT (223kb,D)
[v3] Tue, 14 Nov 2017 20:12:41 GMT (226kb,D)

Link back to: arXiv, form interface, contact.