Skip to main navigation Skip to search Skip to main content

An Improved Multi-task Approach to Pre-trained Model Based MT Quality Estimation

  • Harbin Institute of Technology

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Machine translation (MT) quality estimation (QE) aims to automatically predict the quality of MT outputs without any references. State-of-the-art solutions are mostly fine-tuned with a pre-trained model in a multi-task framework (i.e., joint training sentence-level QE and word-level QE). In this paper, we propose an alternative multi-task framework in which post-editing results are utilized for sentence-level QE over an mBART-based encoder-decoder model. We show that the post-editing sub-task is much more in-formative and the mBART is superior to other pre-trained models. Experiments on WMT2021 English-German and English-Chinese QE datasets showed that the proposed method achieves 1.2%–2.1% improvements in the strong sentence-level QE baseline.

Original languageEnglish
Title of host publicationMachine Translation - 18th China Conference, CCMT 2022, Revised Selected Papers
EditorsTong Xiao, Juan Pino
PublisherSpringer Science and Business Media Deutschland GmbH
Pages106-116
Number of pages11
ISBN (Print)9789811979590
DOIs
StatePublished - 2022
Event18th China Conference on Machine Translation, CCMT 2022 - Lhasa, China
Duration: 6 Aug 202210 Aug 2022

Publication series

NameCommunications in Computer and Information Science
Volume1671 CCIS
ISSN (Print)1865-0929
ISSN (Electronic)1865-0937

Conference

Conference18th China Conference on Machine Translation, CCMT 2022
Country/TerritoryChina
CityLhasa
Period6/08/2210/08/22

Keywords

  • Multitask learning
  • Quality estimation
  • mBART

Fingerprint

Dive into the research topics of 'An Improved Multi-task Approach to Pre-trained Model Based MT Quality Estimation'. Together they form a unique fingerprint.

Cite this