IP Library › Granted Patent US 12,748,934
Granted Patent B2
US 12,748,934 · App. 18/834,798 · Granted Sep 29, 2026

Estimation apparatus, learning apparatus, estimation method, learning method, and program

Inventors: Masaaki Nagata (Tokyo, JP); Yizhen Wei (Ibaraki, JP); Takehito Utsuro (Ibaraki, JP)
Assignees: NTT, Inc.; University of Tsukuba
G06F40/51G06F40/166
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,748,934
App. No.
18/834,798
Granted
Sep 29, 2026
Kind
B2
Abstract

An estimation apparatus includes: an input part configured to receive, as inputs, an extended word alignment between a source sentence and a translated sentence, and translation quality tags between the source sentence and the translated sentence; and an edit tag estimation part configured to estimate an edit tag based on the extended word alignment and the translation quality tags.

Claims (43)

1 . An estimation system comprising:

a processor; and

a memory storing therein a set of instructions which, when executed by the processor, cause the estimation system to:

create an extended word alignment model by training a multilingual model;

store the extended word alignment model in a model database;

receive a source sentence and a machine-translated sentence of the source sentence;

estimate an extended word alignment between the source sentence and the machine-translated sentence of the source sentence using the extended word alignment model, the extended word alignment referring to a function to show, per word, whether or not the machine-translated sentence of the source sentence is a correct translation of the source sentence, and also referring to a correctly or incorrectly machine-translated or aligned word pair;

estimate translation quality tags, with which the source sentence and the machine-translated sentence of the source sentence are labeled per word using a translation quality tag model;

estimate edit tags based on the extended word alignment and the translation quality tags;

divide the edit tags into a replace tag indicating that a first word in the machine-translated sentence of the source sentence is replaced with a correct translation of the first word in the source sentence, an insert tag indicating that a translation of a second word is inserted into the machine-translated sentence of the source sentence and a delete tag indicating that a translation of a third word is deleted from the machine-translated sentence of the source sentence;

label the source sentence with one of the replace tag or the insert tag, label the machine-translated sentence of the source sentence with the delete tag; and

output the labeled source sentence and the labeled machine-translated sentence of the source sentence to a display;

wherein the second word is included in the source sentence, the second word is not included the machine-translated sentence of the source sentence, and

wherein the third word is not included the source sentence, the third word is included the machine-translated sentence of the source sentence.

2 . A computer-implemented estimation method comprising:

creating an extended word alignment model by training a multilingual model;

storing the extended word alignment model in a model database;

receiving a source sentence and a machine-translated sentence of the source sentence;

estimating an extended word alignment between the source sentence and the machine-translated sentence of the source sentence using the extended word alignment model, the extended word alignment referring to a function to show, per word, whether or not the machine-translated sentence of the source sentence is a correct translation of the source sentence, and also referring to a correctly or incorrectly machine-translated or aligned word pair;

estimating translation quality tags, with which the source sentence and the machine-translated sentence of the source sentence are labeled per word using a translation quality tag model;

estimating edit tags based on the extended word alignment and the translation quality tags;

dividing the edit tags into a replace tag indicating that a first word in the machine-translated sentence of the source sentence is replaced with a correct translation of the first word in the source sentence, an insert tag indicating that a translation of a second word is inserted into the machine-translated sentence of the source sentence and a delete tag indicating that a translation of a third word is deleted from the machine-translated sentence of the source sentence;

labeling the source sentence with one of the replace tag or the insert tag, label the machine-translated sentence of the source sentence with the delete tag; and

outputting the labeled source sentence and the labeled machine-translated sentence of the source sentence to a display;

wherein the second word is included in the source sentence, the second word is not included the machine-translated sentence of the source sentence, and

wherein the third word is not included the source sentence, the third word is included the machine-translated sentence of the source sentence.

3 . A computer-readable non-transitory recording medium storing therein a program that, when executed by a computer, causes the computer to perform an estimation method comprising:

creating an extended word alignment model by training a multilingual model;

storing the extended word alignment model in a model database;

receiving a source sentence and a machine-translated sentence of the source sentence;

estimating an extended word alignment between the source sentence and the machine-translated sentence of the source sentence using the extended word alignment model, the extended word alignment referring to a function to show, per word, whether or not the machine-translated sentence of the source sentence is a correct translation of the source sentence, and also referring to a correctly or incorrectly machine-translated or aligned word pair;

estimating translation quality tags, with which the source sentence and the machine-translated sentence of the source sentence are labeled per word using a translation quality tag model;

estimating edit tags based on the extended word alignment and the translation quality tags;

dividing the edit tags into a replace tag indicating that a first word in the machine-translated sentence of the source sentence is replaced with a correct translation of the first word in the source sentence, an insert tag indicating that a translation of a second word is inserted into the machine-translated sentence of the source sentence and a delete tag indicating that a translation of a third word is deleted from the machine-translated sentence of the source sentence;

labeling the source sentence with one of the replace tag or the insert tag, label the machine-translated sentence of the source sentence with the delete tag; and

outputting the labeled source sentence and the labeled machine-translated sentence of the source sentence to a display;

wherein the second word is included in the source sentence, the second word is not included the machine-translated sentence of the source sentence, and

wherein the third word is not included the source sentence, the third word is included the machine-translated sentence of the source sentence.

4 . The estimation system according to claim 1 , wherein the instructions, when executed by the processor, cause the estimation system to:

label a fourth word of the source sentence with one of the replace tag or the insert tag, label a fifth word of the machine-translated sentence of the source sentence with the delete tag; and

wherein the fourth word is labeled with a bad tag of the quality tags using the translation quality tag model,

wherein the fifth word is labeled with the bad tag of the quality tags using the translation quality tag model, and

wherein the bad tag indicates that a word of the source sentence is not translated in the machine-translated sentence of the source sentence, or that a word of the machine-translated sentence of the source sentence is mistranslated.

Assignments (2)
CHANGE OF NAME Recorded Aug 15, 2025
From: NIPPON TELEGRAPH AND TELEPHONE CORPORATION
To: NTT, INC.
Reel/Frame 072490/0664 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jul 31, 2024
From: NAGATA, MASAAKI; WEI, YIZHEN; UTSURO, TAKEHITO
To: NIPPON TELEGRAPH AND TELEPHONE CORPORATION; UNIVERSITY OF TSUKUBA
Reel/Frame 068138/0543 →
Continuity (1)
Related Publication 20250124241A1 · Apr 17, 2025
References Cited (16)
US 6772110B2 · Real · 2004 [cited by examiner]
US 6999916B2 · Lin · 2006 [cited by examiner]
US 10839164B1 · Shorter · 2020 [cited by examiner]
US 11651039B1 · Soubbotin · 2023 [cited by examiner]
US 20010027460A1 · Yamamoto · 2001 [cited by examiner]
US 20090177460A1 · Huang · 2009 [cited by examiner]
US 20100057439A1 · Ideuchi · 2010 [cited by examiner]
US 20150309994A1 · Liu · 2015 [cited by examiner]
US 20160350290A1 · Fujiwara · 2016 [cited by examiner]
US 20180107656A1 · Oda · 2018 [cited by examiner]
US 20190266249A1 · Xu · 2019 [cited by examiner]
US 20200064977A1 · Wu · 2020 [cited by examiner]
US 20220318523A1 · Sheinin · 2022 [cited by examiner]
Hyun Kim et al., “QE BERT: Bilingual BERT using multi-task learning for neural quality estimation”, In Proceedings of the WMT-2019, pp. 85-89, Aug. 1-2, 2019. [cited by applicant]
Fabio Kepler et al., “Unbabel's participation in the WMT19 translation quality estimation shared task”, In Proceedings of the WMT-2019, pp. 78-84, Aug. 1-2, 2019. [cited by applicant]
Masaaki Nagata et al., “A supervised word alignment method based on cross-language span prediction using multilingual BERT”, In Proceedings of EMNLP-2020, pp. 555-565, Nov. 16-20, 2020. [cited by applicant]