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: Context-aware Decoder for Neural Machine Translation using a Target-side Document-Level Language Model

Abstract: Although many context-aware neural machine translation models have been proposed to incorporate contexts in translation, most of those models are trained end-to-end on parallel documents aligned in sentence-level. Because only a few domains (and language pairs) have such document-level parallel data, we cannot perform accurate context-aware translation in most domains. We therefore present a simple method to turn a sentence-level translation model into a context-aware model by incorporating a document-level language model into the decoder. Our context-aware decoder is built upon only a sentence-level parallel corpora and monolingual corpora; thus no document-level parallel data is needed. In a theoretical viewpoint, the core part of this work is the novel representation of contextual information using point-wise mutual information between context and the current sentence. We show the effectiveness of our approach in three language pairs, English to French, English to Russian, and Japanese to English, by evaluation in \textsc{bleu} and contrastive tests for context-aware translation.
Comments: Under Review
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2010.12827 [cs.CL]
  (or arXiv:2010.12827v1 [cs.CL] for this version)

Submission history

From: Amane Sugiyama [view email]
[v1] Sat, 24 Oct 2020 08:06:18 GMT (1110kb,D)

Link back to: arXiv, form interface, contact.