DOCSLIB.ORG

Between Flexibility and Consistency: Joint Generation of Captions And

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Between Flexibility and Consistency: Joint Generation of Captions And Between Flexibility and Consistency: Joint Generation of Captions and Subtitles Alina Karakanta1,2, Marco Gaido1,2, Matteo Negri1, Marco Turchi1 1 Fondazione Bruno Kessler, Via Sommarive 18, Povo, Trento - Italy 2 University of Trento, Italy {akarakanta,mgaido,negri,turchi}@fbk.eu Abstract material but also between each other, for example in the number of blocks (pieces of time-aligned Speech translation (ST) has lately received text) they occupy, their length and segmentation. growing interest for the generation of subtitles Consistency is vital for user experience, for ex- without the need for an intermediate source language transcription and timing (i.e. cap- ample in order to elicit the same reaction among tions). However, the joint generation of source multilingual audiences, or to facilitate the quality captions and target subtitles does not only assurance process in the localisation industry. bring potential output quality advantageswhen Previous work in ST for subtitling has focused the two decoding processes inform each other, on generating interlingual subtitles (Matusov et al., but it is also often required in multilingual sce- 2019; Karakanta et al., 2020a), a) without consid- narios. In this work, we focus on ST models ering the necessity of obtaining captions consis- which generate consistent captions-subtitles in tent with the target subtitles, and b) without ex- terms of structure and lexical content. We further introduce new metrics for evaluating amining whether the joint generation leads to im- subtitling consistency. Our findings show that provements in quality. We hypothesise that knowl- joint decoding leads to increased performance edge sharing between the tasks of transcription and consistency between the generated cap- and translation could lead to such improvements. tions and subtitles while still allowing for suf- Moreover, joint generation with a single system ficient flexibility to produce subtitles conform- can avoid the maintenance of two different models, ing to language-specific needs and norms. increase efficiency, and in turn speed up the locali- 1 Introduction sation process. Lastly, if joint generation improves consistency, joint models could increase automa- New trends in media localisation call for the rapid tion in subtitling applications where consistency is generation of subtitles for vast amounts of au- a desideratum. diovisual content. Speech translation, and espe- In this work, we address these issues for the cially direct approaches (B´erard et al., 2016; Ba- first time, by jointly generating both captions and har et al., 2019), have recently shown promising subtitles for the same audio source. We experi- results with high efficiency because they do not ment with the following models: 1) Shared Direct arXiv:2107.06246v1 [cs.CL] 13 Jul 2021 require a transcription (manual or automatic) of (Weiss et al., 2017), where the speech encoder is the source speech but generate the target language shared between the transcription and the transla- subtitles directly from the audio. However, obtain- tion decoder, 2) Two-Stage (Kano et al., 2017), ing the intralingual subtitles (hereafter “captions”) where the transcript decoder states are passed to is necessary for a range of applications, while in the translation decoder, and 3) Triangle (Anasta- some settings captions need to be displayed along sopoulos and Chiang, 2018), which extends the with the target language subtitles. Such “bilingual two-stage by adding a second attention mecha- subtitles” are useful in multilingual online meet- nism to the translation decoder which attends to ings, in countries with multiple official languages, encoded speech inputs. We compare these mod- or for language learners and audiences with dif- els with the established approaches in ST for sub- ferent accessibility needs. In those cases, captions titling: an independent direct ST model and a cas- and subtitles should not only be consistent with the cade (ASR+MT) model. Moreover, we extend the visual and acoustic dimension of the audiovisual evaluation beyond the usual metrics used to assess transcription and translation quality (respectively been common in countries with more than one offi- WER and BLEU), by also evaluating the form and cial languages, such as Belgium and Finland (Got- consistency of the generated subtitles. tlieb, 2004). Recently, however, captions along Sperber et al. (2020) introduced several lexical with subtitles have been employed in other coun- and surface metrics to measure consistency of ST tries to attract wider audiences, e.g. in Mainland outputs, but they were only applied to standard, China English captions are displayed along with non-subtitle, texts. Subtitles, however, are a partic- Mandarin subtitles. Interestingly, despite doubling ular type of text structured in blocks which accom- the amount of text that appears on the screen and pany the action on screen. Therefore, we propose the high redundancy, it has been shown that bilin- to measure their consistency by taking advantage gual subtitles do not significantly increase users’ of this structure and introduce metrics able to re- cognitive load (Liao et al., 2020). ward subtitles that share similar structure and con- One group which undoubtedly benefits from the tent. parallel presence of captions and subtitles are lan- Our contributions can be summarised as fol- guage learners. Captions have been found to in- lows: crease learners’ L2 vocabulary (Sydorenko, 2010) and improve listening comprehension (Guichon • We employ ST to directly generate both cap- and McLornan, 2008). Subtitles in the learn- tions and subtitles without the need for hu- ers’ native language (L1) are an indispensable man pre-processing (transcription, segmenta- tool for comprehension and access to information, tion). especially for beginners. In bilingual subtitles, the captions support learners in understanding the • We propose new, task-specific metrics to eval- speech and acquiring terminology, while subtitles uate subtitling consistency, a challenging and serve as a dictionary, facilitating bilingual map- understudied problem in subtitling evalua- ping (Garc´ıa, 2017). Consistency is particularly tion. important for bilingual subtitles. Terms should fall in the same block and in similar positions. More- • We show increased performance and consis- over, similar length and equal number of lines can tency between the generated captions and prevent long distance saccades, assisting in spot- subtitles compared to independent decoding, ting the necessary information in the two language while preserving adequate conformity to sub- versions. Several subtitling tools have recently al- titling constraints. lowed for combining captions and subtitles on the same video (e.g. Dualsub1) and bilingual subti- 2 Background tles can be obtained for Youtube videos2 and TED 3 2.1 Bilingual subtitles Talks. Another aspect where consistency between cap- New life conditions maximised the time spent tions and subtitles is present is in subtitling tem- in front of screens, transforming today’s medias- plates. A subtitling template is a source lan- cape in a complex puzzle of new actors, voices guage/English version of a subtitle file already and workflows. Face-to-face meetings and con- segmented and containing timestamps, which is ferences moved online, opening up participation used to directly translate the text in target lan- for global audiences. In these new settings multi- guages while preserving the same structure (Cin- linguality and versatility are dominant, and mani- tas and Remael, 2007; Georgakopoulou, 2019; fested in business meetings with international part- Netflix, 2021). This process reduces the cost, turn- ners, conferences with world-wide coverage, mul- around times and effort needed to create a sepa- tilingual classrooms and audiences with mixed ac- rate timed version for each language, and facili- cessibility needs. Given these growing needs for tates quality assurance since errors can be spotted providing inclusive and equal access to audiovi- across the same blocks (Nikoli´c, 2015). These ben- sual material for a multifaceted audience spectrum, efits motivated our work towards simultaneously efficiently obtaining high-quality captions and sub- titles is becoming increasingly relevant. 1https://www.dualsub.xyz/ Traditionally, displaying subtitles in two lan- 2https://www.watch-listen-read.com/ guages in parallel (bilingual or dual subtitles) has 3https://amara.org/en/teams/ted/ generating two language versions with the maxi- et al. (2020) evaluated these methods with the mum consistency, where the caption file can fur- focus of jointly producing consistent source and ther serve as a template for multilingual localisa- target language texts. Their underlying intuition tion. This paper is a first step towards maximising is that, since in cascade solutions the translation automation for the generation of high-quality mul- is derived from the transcript, cascades should tiple language/accessibility subtitle versions. achieve higher consistency than direct solutions. Their results, however, showed that triangle mod- 2.2 MT and ST for subtitling els achieve the highest consistency among the ar- Subtitling has long sparked the interest of the Ma- chitectures tested (considerably better than that chine Translation (MT) community as a challeng- of cascade systems) and have competitive per- ing type of translation. Most works employing MT formance in terms of translation quality. Direct for subtitling stem from the statistical
Load more

Recommended publications
  • Can You Give Me Another Word for Hyperbaric?: Improving Speech
    “CAN YOU GIVE ME ANOTHER WORD FOR HYPERBARIC?”: IMPROVING SPEECH TRANSLATION USING TARGETED CLARIFICATION QUESTIONS Necip Fazil Ayan1, Arindam Mandal1, Michael Frandsen1, Jing Zheng1, Peter Blasco1, Andreas Kathol1 Fred´ eric´ Bechet´ 2, Benoit Favre2, Alex Marin3, Tom Kwiatkowski3, Mari Ostendorf3 Luke Zettlemoyer3, Philipp Salletmayr5∗, Julia Hirschberg4, Svetlana Stoyanchev4 1 SRI International, Menlo Park, USA 2 Aix-Marseille Universite,´ Marseille, France 3 University of Washington, Seattle, USA 4 Columbia University, New York, USA 5 Graz Institute of Technology, Austria ABSTRACT Our previous work on speech-to-speech translation systems has We present a novel approach for improving communication success shown that there are seven primary sources of errors in translation: between users of speech-to-speech translation systems by automat- • ASR named entity OOVs: Hi, my name is Colonel Zigman. ically detecting errors in the output of automatic speech recogni- • ASR non-named entity OOVs: I want some pristine plates. tion (ASR) and statistical machine translation (SMT) systems. Our approach initiates system-driven targeted clarification about errorful • Mispronunciations: I want to collect some +de-MOG-raf-ees regions in user input and repairs them given user responses. Our about your family? (demographics) system has been evaluated by unbiased subjects in live mode, and • Homophones: Do you have any patients to try this medica- results show improved success of communication between users of tion? (patients vs. patience) the system. • MT OOVs: Where is your father-in-law? Index Terms— Speech translation, error detection, error correc- • Word sense ambiguity: How many men are in your com- tion, spoken dialog systems. pany? (organization vs.
    [Show full text]
  • A Survey of Voice Translation Methodologies - Acoustic Dialect Decoder
    International Conference On Information Communication & Embedded Systems (ICICES-2016) A SURVEY OF VOICE TRANSLATION METHODOLOGIES - ACOUSTIC DIALECT DECODER Hans Krupakar Keerthika Rajvel Dept. of Computer Science and Engineering Dept. of Computer Science and Engineering SSN College Of Engineering SSN College of Engineering E-mail: [email protected] E-mail: [email protected] Bharathi B Angel Deborah S Vallidevi Krishnamurthy Dept. of Computer Science Dept. of Computer Science Dept. of Computer Science and Engineering and Engineering and Engineering SSN College Of Engineering SSN College Of Engineering SSN College Of Engineering E-mail: [email protected] E-mail: [email protected] E-mail: [email protected] Abstract— Speech Translation has always been about giving in order to make the process of communication amongst source text/audio input and waiting for system to give humans better, easier and efficient. However, the existing translated output in desired form. In this paper, we present the methods, including Google voice translators, typically Acoustic Dialect Decoder (ADD) – a voice to voice ear-piece handle the process of translation in a non-automated manner. translation device. We introduce and survey the recent advances made in the field of Speech Engineering, to employ in This makes the process of translation of word(s) and/or the ADD, particularly focusing on the three major processing sentences from one language to another, slower and more steps of Recognition, Translation and Synthesis. We tackle the tedious. We wish to make that process automatic – have a problem of machine understanding of natural language by device do what a human translator does, inside our ears.
    [Show full text]
  • Learning Speech Translation from Interpretation
    Learning Speech Translation from Interpretation zur Erlangung des akademischen Grades eines Doktors der Ingenieurwissenschaften von der Fakult¨atf¨urInformatik Karlsruher Institut f¨urTechnologie (KIT) genehmigte Dissertation von Matthias Paulik aus Karlsruhe Tag der m¨undlichen Pr¨ufung: 21. Mai 2010 Erster Gutachter: Prof. Dr. Alexander Waibel Zweiter Gutachter: Prof. Dr. Tanja Schultz Ich erkl¨arehiermit, dass ich die vorliegende Arbeit selbst¨andig verfasst und keine anderen als die angegebenen Quellen und Hilfsmittel verwendet habe sowie dass ich die w¨ortlich oder inhaltlich ¨ubernommenen Stellen als solche kenntlich gemacht habe und die Satzung des KIT, ehem. Universit¨atKarlsruhe (TH), zur Sicherung guter wissenschaftlicher Praxis in der jeweils g¨ultigenFassung beachtet habe. Karlsruhe, den 21. Mai 2010 Matthias Paulik Abstract The basic objective of this thesis is to examine the extent to which automatic speech translation can benefit from an often available but ignored resource, namely human interpreter speech. The main con- tribution of this thesis is a novel approach to speech translation development, which makes use of that resource. The performance of the statistical models employed in modern speech translation systems depends heavily on the availability of vast amounts of training data. State-of-the-art systems are typically trained on: (1) hundreds, sometimes thousands of hours of manually transcribed speech audio; (2) bi-lingual, sentence-aligned text corpora of man- ual translations, often comprising tens of millions of words; and (3) monolingual text corpora, often comprising hundreds of millions of words. The acquisition of such enormous data resources is highly time-consuming and expensive, rendering the development of deploy- able speech translation systems prohibitive to all but a handful of eco- nomically or politically viable languages.
    [Show full text]
  • Is 42 the Answer to Everything in Subtitling-Oriented Speech Translation?
    Is 42 the Answer to Everything in Subtitling-oriented Speech Translation? Alina Karakanta Matteo Negri Marco Turchi Fondazione Bruno Kessler Fondazione Bruno Kessler Fondazione Bruno Kessler University of Trento Trento - Italy Trento - Italy Trento - Italy [email protected] [email protected] [email protected] Abstract content, still rely heavily on human effort. In a typi- cal multilingual subtitling workflow, a subtitler first Subtitling is becoming increasingly important creates a subtitle template (Georgakopoulou, 2019) for disseminating information, given the enor- by transcribing the source language audio, timing mous amounts of audiovisual content becom- and adapting the text to create proper subtitles in ing available daily. Although Neural Machine Translation (NMT) can speed up the process the source language. These source language subti- of translating audiovisual content, large man- tles (also called captions) are already compressed ual effort is still required for transcribing the and segmented to respect the subtitling constraints source language, and for spotting and seg- of length, reading speed and proper segmentation menting the text into proper subtitles. Cre- (Cintas and Remael, 2007; Karakanta et al., 2019). ating proper subtitles in terms of timing and In this way, the work of an NMT system is already segmentation highly depends on information simplified, since it only needs to translate match- present in the audio (utterance duration, natu- ing the length of the source text (Matusov et al., ral pauses). In this work, we explore two meth- ods for applying Speech Translation (ST) to 2019; Lakew et al., 2019). However, the essence subtitling: a) a direct end-to-end and b) a clas- of a good subtitle goes beyond matching a prede- sical cascade approach.
    [Show full text]
  • Speech Translation Into Pakistan Sign Language
    Master Thesis Computer Science Thesis No: MCS-2010-24 2012 Speech Translation into Pakistan Sign Language Speech Translation into Pakistan Sign AhmedLanguage Abdul Haseeb Asim Illyas Speech Translation into Pakistan Sign LanguageAsim Ilyas School of Computing Blekinge Institute of Technology SE – 371 79 Karlskrona Sweden This thesis is submitted to the School of Computing at Blekinge Institute of Technology in partial fulfillment of the requirements for the degree of Master of Science in Computer Science. The thesis is equivalent to 20 weeks of full time studies. Contact Information: Authors: Ahmed Abdul Haseeb E-mail: [email protected] Asim Ilyas E-mail: [email protected] University advisor: Prof. Sara Eriksén School of Computing, Blekinge Institute of Technology Internet : www.bth.se/com SE – 371 79 Karlskrona Phone : +46 457 38 50 00 Sweden Fax : + 46 457 271 25 ABSTRACT Context: Communication is a primary human need and language is the medium for this. Most people have the ability to listen and speak and they use different languages like Swedish, Urdu and English etc. to communicate. Hearing impaired people use signs to communicate. Pakistan Sign Language (PSL) is the preferred language of the deaf in Pakistan. Currently, human PSL interpreters are required to facilitate communication between the deaf and hearing; they are not always available, which means that communication among the deaf and other people may be impaired or nonexistent. In this situation, a system with voice recognition as an input and PSL as an output will be highly helpful. Objectives: As part of this thesis, we explore challenges faced by deaf people in everyday life while interacting with unimpaired.
    [Show full text]
  • Implementing Machine Translation and Post-Editing to the Translation of Wildlife Documentaries Through Voice-Over and Off-Screen Dubbing
    ADVERTIMENT. Lʼaccés als continguts dʼaquesta tesi queda condicionat a lʼacceptació de les condicions dʼús establertes per la següent llicència Creative Commons: http://cat.creativecommons.org/?page_id=184 ADVERTENCIA. El acceso a los contenidos de esta tesis queda condicionado a la aceptación de las condiciones de uso establecidas por la siguiente licencia Creative Commons: http://es.creativecommons.org/blog/licencias/ WARNING. The access to the contents of this doctoral thesis it is limited to the acceptance of the use conditions set by the following Creative Commons license: https://creativecommons.org/licenses/?lang=en Universitat Autònoma de Barcelona Departament de Traducció i d’Interpretació i d’Estudis de l’Asia Oriental Doctorat en Traducció i Estudis Interculturals Implementing Machine Translation and Post-Editing to the Translation of Wildlife Documentaries through Voice-over and Off-screen Dubbing A Research on Effort and Quality PhD dissertation presented by: Carla Ortiz Boix Supervised and tutorized by: Dr. Anna Matamala 2016 A la meva família: als que hi són, als que no, i als que només hi són a mitges. Acknowledgments The road to finishing this PhD has not been easy and it would not have been accomplished without the priceless support of many: First of all, I want to thank my supervisor, Dr. Anna Matamala, for all her hard work. It has not been an easy road and sometimes I would have lost the right path if she had not been there to support, encourage, and challenge me. The PhD would not have come out the way it has without you. On a professional level, I also want to thank Dr.
    [Show full text]
  • Breeding Gender-Aware Direct Speech Translation Systems
    Breeding Gender-aware Direct Speech Translation Systems Marco Gaido1,2 y, Beatrice Savoldi2 y, Luisa Bentivogli1, Matteo Negri1, Marco Turchi1 1Fondazione Bruno Kessler, Trento, Italy 2University of Trento, Italy fmgaido,bentivo,negri,[email protected],[email protected] Abstract In automatic speech translation (ST), traditional cascade approaches involving separate transcrip- tion and translation steps are giving ground to increasingly competitive and more robust direct solutions. In particular, by translating speech audio data without intermediate transcription, di- rect ST models are able to leverage and preserve essential information present in the input (e.g. speaker’s vocal characteristics) that is otherwise lost in the cascade framework. Although such ability proved to be useful for gender translation, direct ST is nonetheless affected by gender bias just like its cascade counterpart, as well as machine translation and numerous other natural language processing applications. Moreover, direct ST systems that exclusively rely on vocal biometric features as a gender cue can be unsuitable and potentially harmful for certain users. Going beyond speech signals, in this paper we compare different approaches to inform direct ST models about the speaker’s gender and test their ability to handle gender translation from English into Italian and French. To this aim, we manually annotated large datasets with speakers’ gen- der information and used them for experiments reflecting different possible real-world scenarios. Our results show that gender-aware direct ST solutions can significantly outperform strong – but gender-unaware – direct ST models. In particular, the translation of gender-marked words can increase up to 30 points in accuracy while preserving overall translation quality.
    [Show full text]
  • TAUS Speech-To-Speech Translation Technology Report March 2017
    TAUS Speech-to-Speech Translation Technology Report March 2017 1 Authors: Mark Seligman, Alex Waibel, and Andrew Joscelyne Contributor: Anne Stoker COPYRIGHT © TAUS 2017 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system of any nature, or transmit- ted or made available in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior written permission of TAUS. TAUS will pursue copyright infringements. In spite of careful preparation and editing, this publication may contain errors and imperfections. Au- thors, editors, and TAUS do not accept responsibility for the consequences that may result thereof. Design: Anne-Maj van der Meer Published by TAUS BV, De Rijp, The Netherlands For further information, please email [email protected] 2 Table of Contents Introduction 4 A Note on Terminology 5 Past 6 Orientation: Speech Translation Issues 6 Present 17 Interviews 19 Future 42 Development of Current Trends 43 New Technology: The Neural Paradigm 48 References 51 Appendix: Survey Results 55 About the Authors 56 3 Introduction The dream of automatic speech-to-speech devices enabling on-the-spot exchanges using translation (S2ST), like that of automated trans- voice or text via smartphones – has helped lation in general, goes back to the origins of promote the vision of natural communication computing in the 1950s. Portable speech trans- on a globally connected planet: the ability to lation devices have been variously imagined speak to someone (or to a robot/chatbot) in as Star Trek’s “universal translator” to negoti- your language and be immediately understood ate extraterrestrial tongues, Douglas Adams’ in a foreign language.
    [Show full text]
  • Arxiv:2006.01080V1 [Cs.CL] 1 Jun 2020
    Is 42 the Answer to Everything in Subtitling-oriented Speech Translation? Alina Karakanta Matteo Negri Marco Turchi Fondazione Bruno Kessler Fondazione Bruno Kessler Fondazione Bruno Kessler University of Trento Trento - Italy Trento - Italy Trento - Italy [email protected] [email protected] [email protected] Abstract content, still rely heavily on human effort. In a typi- cal multilingual subtitling workflow, a subtitler first Subtitling is becoming increasingly important creates a subtitle template (Georgakopoulou, 2019) for disseminating information, given the enor- by transcribing the source language audio, timing mous amounts of audiovisual content becom- and adapting the text to create proper subtitles in ing available daily. Although Neural Machine Translation (NMT) can speed up the process the source language. These source language subti- of translating audiovisual content, large man- tles (also called captions) are already compressed ual effort is still required for transcribing the and segmented to respect the subtitling constraints source language, and for spotting and seg- of length, reading speed and proper segmentation menting the text into proper subtitles. Cre- (Cintas and Remael, 2007; Karakanta et al., 2019). ating proper subtitles in terms of timing and In this way, the work of an NMT system is already segmentation highly depends on information simplified, since it only needs to translate match- present in the audio (utterance duration, natu- ing the length of the source text (Matusov et al., ral pauses). In this work, we explore two meth- ods for applying Speech Translation (ST) to 2019; Lakew et al., 2019). However, the essence subtitling: a) a direct end-to-end and b) a clas- of a good subtitle goes beyond matching a prede- sical cascade approach.
    [Show full text]
  • Multilingual Named Entity Extraction and Translation from Text and Speech
    MULTILINGUAL NAMED ENTITY EXTRACTION AND TRANSLATION FROM TEXT AND SPEECH Fei Huang December 2005 CMU-LTI-06-001 Language Technologies Institute School of Computer Science Carnegie Mellon University 5000 Forbes Ave., Pittsburgh, PA 15213 www.lti.cs.cmu.edu THESIS COMMITTEE Alexander Waibel (Chair) Stephan Vogel Alon Lavie Tanja Schultz Kevin Knight (USC/ISI) Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy °c 2005 Fei Huang 1 Acknowledgements I want to thank this age, the age of globalization, especially its posi- tive effect on China. When I was 10 years old, I could not imagine I would visit another country tens of thousands of miles away, living there for more than six years for my doctoral study, and having the opportunity of appreciating different cultures. I hope the globaliza- tion will continue to shape everyone’s life in a positive manner. I owe a lot to my thesis committee members. I greatly benefited from the scientific vision of my advisor, Alex Waibel. He often pointed out new directions in my research, helped me develop in- dependent research ability and challenged my over-prudence. ”Be BOLD!” Yes I will keep this as my lifelong reminder. I also appre- ciate the constant support from my direct supervisor, mentor and friend, Stephan Vogel. He has always been available for both scien- tific and causal discussions. I gained a lot from his rich experience in machine translation. Working with him has been fun, especially during the DARPA machine translation evaluations! Alon Lavie and Kevin Knight helped me both scientifically and professionally.
    [Show full text]
  • Integration of Speech Recognition and Machine Translation in Computer-Assisted Translation Shahram Khadivi and Hermann Ney, Senior Member, IEEE
    IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 16, NO. 8, NOVEMBER 2008 1551 Integration of Speech Recognition and Machine Translation in Computer-Assisted Translation Shahram Khadivi and Hermann Ney, Senior Member, IEEE Abstract—Parallel integration of automatic speech recognition (ASR) models and statistical machine translation (MT) models is an unexplored research area in comparison to the large amount of works done on integrating them in series, i.e., speech-to-speech translation. Parallel integration of these models is possible when we have access to the speech of a target language text and to its corre- sponding source language text, like a computer-assisted translation system. To our knowledge, only a few methods for integrating ASR models with MT models in parallel have been studied. In this paper, we systematically study a number of different translation models in the context of the -best list rescoring. As an alternative to the -best list rescoring, we use ASR word graphs in order to arrive at a tighter integration of ASR and MT models. The experiments are carried out on two tasks: English-to-German with an ASR vo- cabulary size of 17 K words, and Spanish-to-English with an ASR Fig. 1. Schematic of automatic text dictation in computer-assisted translation. vocabulary of 58 K words. For the best method, the MT models re- duce the ASR word error rate by a relative of 18% and 29% on the 17 K and the 58 K tasks, respectively. language speech is a human-produced translation of the source Index Terms—Computer-assisted translation (CAT), speech language text.
    [Show full text]
  • Automatic Post-Editing for Machine Translation Rajen Chatterjee
    DEPARTMENT OF INFORMATION ENGINEERING AND COMPUTER SCIENCE ICT International Doctoral School Automatic Post-Editing for Machine Translation Rajen Chatterjee Advisor Dr. Marco Turchi Fondazione Bruno Kessler Co-Advisor arXiv:1910.08592v1 [cs.CL] 18 Oct 2019 Matteo Negri Fondazione Bruno Kessler October 2019 Abstract Automatic Post-Editing (APE) aims to correct systematic errors in a ma- chine translated text. This is primarily useful when the machine translation (MT) system is not accessible for improvement, leaving APE as a viable op- tion to improve translation quality as a downstream task - which is the focus of this thesis. This field has received less attention compared to MT due to several reasons, which include: the limited availability of data to perform a sound research, contrasting views reported by different researchers about the effectiveness of APE, and limited attention from the industry to use APE in current production pipelines. In this thesis, we perform a thorough investigation of APE as a down- stream task in order to: i) understand its potential to improve translation quality; ii) advance the core technology - starting from classical methods to recent deep-learning based solutions; iii) cope with limited and sparse data; iv) better leverage multiple input sources; v) mitigate the task-specific problem of over-correction; vi) enhance neural decoding to leverage external knowledge; and vii) establish an online learning framework to handle data diversity in real-time. All the above contributions are discussed across several chapters, and most of them are evaluated in the APE shared task organized each year at the Conference on Machine Translation. Our efforts in improving the technology resulted in the best system at the 2017 APE shared task, and our work on online learning received a distinguished paper award at the Italian Conference on Computational Linguistics.
    [Show full text]