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


Current browse context:


Change to browse by:


References & Citations

DBLP - CS Bibliography


(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: Towards Linear Time Neural Machine Translation with Capsule Networks

Abstract: In this study, we first investigate a novel capsule network with dynamic routing for linear time Neural Machine Translation (NMT), referred as \textsc{CapsNMT}. \textsc{CapsNMT} uses an aggregation mechanism to map the source sentence into a matrix with pre-determined size, and then applys a deep LSTM network to decode the target sequence from the source representation. Unlike the previous work \cite{sutskever2014sequence} to store the source sentence with a passive and bottom-up way, the dynamic routing policy encodes the source sentence with an iterative process to decide the credit attribution between nodes from lower and higher layers. \textsc{CapsNMT} has two core properties: it runs in time that is linear in the length of the sequences and provides a more flexible way to select, represent and aggregates the part-whole information of the source sentence. On WMT14 English-German task and a larger WMT14 English-French task, \textsc{CapsNMT} achieves comparable results with the state-of-the-art NMT systems. To the best of our knowledge, this is the first work that capsule networks have been empirically investigated for sequence to sequence problems.
Comments: Accepted as EMNLP2019
Subjects: Computation and Language (cs.CL)
DOI: 10.18653/v1/D19-1074
Cite as: arXiv:1811.00287 [cs.CL]
  (or arXiv:1811.00287v4 [cs.CL] for this version)

Submission history

From: Mingxuan Wang [view email]
[v1] Thu, 1 Nov 2018 09:37:48 GMT (300kb,D)
[v2] Thu, 29 Aug 2019 13:32:33 GMT (191kb,D)
[v3] Tue, 14 Jul 2020 06:11:14 GMT (191kb,D)
[v4] Mon, 12 Oct 2020 03:41:02 GMT (191kb,D)

Link back to: arXiv, form interface, contact.