MtJaen

Japanese English Machine Translation

This page describes the Jaen MT system, an MT system based on the [wiki:LogonTop LOGON architecture], using [wiki:JacyTop Jacy] for the source language analysis and the [wiki:ErgTop ERG] for the target language generation. In order to degrade gracefully in the presence of input we cannot yet translate, we are also working on a backup [wiki:MtJaenSmt Statistical MT system] based on [http://www.statmt.org/moses/ Moses].

Jaen is the elder sibling of [wiki:NoJa Noja], the Norwegian-Japanese MT system.

The first version of Jaen is described in:

• Francis Bond, Stephan Oepen, Melanie Siegel, Ann Copestake, and Dan Flickinger (2005)
  • [http://www2.nict.go.jp/x/x161/en/member/bond/pubs/2005-summit-osmt.pdf Open source machine translation with DELPH-IN]. In Open-Source Machine Translation: Workshop at MT Summit X, pp 15--22, Phuket.

Tanaka Corpus Development Data

Change	Date	Parse Coverage	Transfer Coverage	Generation Coverage	End-to-End Coverage	BLEU	Oracle
Initial run	06/05	2779 / 4500 (61.76%)	549 / 2779 (19.76%)	383 / 549 (69.76%)	383 / 4500 (8.51%)	0.1433	0.2234
Fixed _eba_c_rel etc.	06/07	2779 / 4500 (61.76%)	662 / 2779 (23.82%)	478 / 662 (72.21%)	478 / 4500 (10.62%)	0.1436	0.2302
Rel name debugging + handcrafted rules	06/09	2779 / 4500 (61.76%)	679 / 2779 (24.43%)	487 / 679 (71.72%)	487 / 4500 (10.82%)	0.1470	0.2335
Generic entries w/o にる,だける	06/14	3014 / 4500 (66.98%)	691 / 3014 (22.93%)	491 / 691 (71.06%)	491 / 4500 (10.91%)	0.1404	0.2300
New vn handling	06/14	3014 / 4500 (66.98%)	703 / 3014 (23.32%)	500 / 703 (71.12%)	500 / 4500 (11.11%)	0.1369	0.2264
Generic Edict rules	06/15	3014 / 4500 (66.98%)	722 / 3014 (23.95%)	520 / 722 (72.02%)	520 / 4500 (11.56%)	0.1337	0.2231