Teksti stiiliülekanne ühekeelse masintõlke abil
Date
2019
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Teksti stiiliülekande ülesanne on muuta teksti stiilseid omadusi, säilitades samal ajal selle tähenduse ja soravuse. Seda saab vaadata järjendite teisenduse ülesandena, kuid otseselt märgendatud paralleelandmete nappus muudab selle enamiku juhtumite jaoks võimatuks. Käesolevas töös me kirjeldame stiiliülekande lähenemist, mis põhineb zero-shot masintõlke ideel. Meie pakutud mudel sooritab stiiliülekande neuromasintõlke abil, ilma nõudmata paralleelseid stiili-kohandatud tekste, toetudes hoopis regulaarse keele paralleelsetele andmetele. Meetodit saab rakendada mitmele keelele ühes ja samas treenitud mudelis. Meie kirjeldame pakutud metodoloogiat, oma katseid sellega ning esitame põhjaliku automaatse ja käsitsi teostatava hindamise, kus võrdleme seda juhendatud masinõppe põhise baasmudeliga. Me näitame, et meie mudeli tulemuste hinnangud ületavad juhendatud baaslähenemist mitmes aspektis vastavalt inimeste arvamustele ning on usaldusväärsed mitme stiiliülekande aspekti suhtes.
Text style transfer is the task of altering stylistic features of a text while preserving its meaning and fluency. It can be viewed as a sequence-to-sequence transformation task, but the scarcity of directly annotated parallel data makes it unfeasible for most settings. We propose an approach to style transfer that builds on the idea of zero-shot machine translation. It performs style transfer within a neural machine translation model, without requiring any parallel style-adapted texts, relying instead only on regular language-parallel data. The method is applicable to multiple languages within a single model. We outline the method, describe our experiments with it, and, finally, present a thorough automatic and manual evaluation of the approach in comparison to a baseline and independently, showing that our zero-shot model outperforms the supervised baseline on several aspects according to human judgments, and is reliable for a number of style transfer aspects, while not depending on annotated data.
Text style transfer is the task of altering stylistic features of a text while preserving its meaning and fluency. It can be viewed as a sequence-to-sequence transformation task, but the scarcity of directly annotated parallel data makes it unfeasible for most settings. We propose an approach to style transfer that builds on the idea of zero-shot machine translation. It performs style transfer within a neural machine translation model, without requiring any parallel style-adapted texts, relying instead only on regular language-parallel data. The method is applicable to multiple languages within a single model. We outline the method, describe our experiments with it, and, finally, present a thorough automatic and manual evaluation of the approach in comparison to a baseline and independently, showing that our zero-shot model outperforms the supervised baseline on several aspects according to human judgments, and is reliable for a number of style transfer aspects, while not depending on annotated data.