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

Download:

Current browse context:

cs.CL

Change to browse by:

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 > Computation and Language

Title: DISCO : efficient unsupervised decoding for discrete natural language problems via convex relaxation

Abstract: In this paper we study test time decoding; an ubiquitous step in almost all sequential text generation task spanning across a wide array of natural language processing (NLP) problems. Our main contribution is to develop a continuous relaxation framework for the combinatorial NP-hard decoding problem and propose Disco - an efficient algorithm based on standard first order gradient based. We provide tight analysis and show that our proposed algorithm linearly converges to within $\epsilon$ neighborhood of the optima. Finally, we perform preliminary experiments on the task of adversarial text generation and show superior performance of Disco over several popular decoding approaches.
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Cite as: arXiv:2107.05380 [cs.CL]
  (or arXiv:2107.05380v2 [cs.CL] for this version)

Submission history

From: Anish Acharya [view email]
[v1] Wed, 7 Jul 2021 00:40:25 GMT (206kb,D)
[v2] Tue, 13 Jul 2021 20:34:37 GMT (206kb,D)

Link back to: arXiv, form interface, contact.