Improve Translation Yakov Kronrod 1 Chang Hu 2 Olivia Buzek 2 Alex Quinn 2 University of Maryland 1 Dept of Linguistics 2 Dept of Computer Science UMD HCI Lab csumdedu ID: 344625
Download Presentation The PPT/PDF document "Using Monolingual Crowds to" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Using Monolingual Crowds to
Improve Translation
Yakov Kronrod
1, Chang Hu2, Olivia Buzek2, Alex Quinn2 University of Maryland (1) Dept. of Linguistics (2) Dept. of Computer Science
UMD
HCI Lab (cs.umd.edu/hcil) Comp. Ling. and Info. Processing Lab (umiacs.umd.edu/labs/clip) MonoTrans Project Page: (http://www.cs.umd.edu/hcil/monotrans)
1
ParaTrans
: Targeted Paraphrasing
Monolingual speakers
improving translation quality
Human participants perform the following tasks:Identify mistranslated spans in target languageCreate paraphrases for spans in source language responsible for the target error spans.Produces gains of 0.82 and 1.32 for adequacy and fluency on a 5-point scaleImproves 65.3% of sentences examinedYields increases of +0.9 BLEU to +3.8 BLEU, depending on size of error span, in an oracle study.
Monolingual Human Participation
Quality
Affordability
Space of Available Translation Options
Online system for collaborative monolingual translation
Asynchronous edits, short
tasks, minimal context
repeat …
Target Language
Source Language
MT
Sentence
Sentence
1
Vote
2
Identify error spans
3
Create new
translation
candidates
1
Vote
2
Paraphrase
3
Explain errors
Crowd
Tasks:
New candidate
1
2
3
MT and
word alignment
…
Crowd
Tasks:
The Bigger Picture: MonoTrans Features Results
Translations and Human Judgments
Original Google Translate output (GT), with translations produced by the targeted paraphrase translation process (TP), selected to show a range from strong to weak improvements in fluency. Fluency and adequacy (preservation of meaning) were rated on a 5-point scale by independent judges not familiar with this research.
Paraphrases from the Crowd
68% of sentences translated perfectly with
no
bilingual humans