Interactive translation with neural models based on the use of mouse actions and confidence measures
[EN] The state of the art in machine translation is not good enough to be able to guarantee high quality translations at all times. In some fields, it is necessary to always obtain error-free translations, and the tendency is to use the help of a professional to correct them, making use of processes...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2020 |
| País: | España |
| Institución: | Universitat Politècnica de València (UPV) |
| Repositorio: | RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia |
| Idioma: | inglés |
| OAI Identifier: | oai:riunet.upv.es:10251/151718 |
| Acceso en línea: | https://riunet.upv.es/handle/10251/151718 |
| Access Level: | acceso abierto |
| Palabra clave: | Traducción Automática Traducción automática interactiva Traducción Neuronal Multimodalidad Acciónes de Ratón Medidas de Confianza Machine Translation Interactive Machine Translation Neural Machine Translation Multimodality. Mouse Actions Confidence Measures LENGUAJES Y SISTEMAS INFORMATICOS Máster Universitario en Inteligencia Artificial, Reconocimiento de Formas e Imagen Digital-Màster Universitari en Intel·ligència Artificial, Reconeixement de Formes i Imatge Digital |
| Sumario: | [EN] The state of the art in machine translation is not good enough to be able to guarantee high quality translations at all times. In some fields, it is necessary to always obtain error-free translations, and the tendency is to use the help of a professional to correct them, making use of processes such as post-editing or interactive translation. The usual way of human-machine interaction is through the use of the keyboard and mouse. The user positions the cursor in front of an incorrect word, type the correction, and the system provides a new suffix. In this master’s thesis, we are going to develop two extensions to this system: the integration of mouse actions, and the use of confidence measures. These extensions have already been developed previously for statistical models, but have not yet been developed and tested in neural models. Once we develop the extensions in an interactive neural translation system, we will compare the results of the experiments to see the improvement obtained. In the first proposal, we introduce the use of mouse actions as the only input information to the system to correct the translations. The position of the mouse when correcting a word offers enough information to the system to be able to generate a new suffix without typing in the correction. The translation is correct from the beginning to the position where the user has moved the cursor, also indicating that the next word is incorrect. Furthermore, if the generated suffix is incorrect again, a new correction can be requested, providing the system with the extra information that the next word is incorrect again. Finally, an approach has also been developed where the user can move the cursor in the middle of the words, correcting at the character level. In the second proposal, we reduce the effort that a translator has to make, by reducing the number of sentences and words that they have to correct. In conventional interactive translation systems, the human translator has to check every sentence and every word. In this extension, the system provides us with an estimation of how correct it thinks the translated words are, and the user only has to check those that do not exceed a certain threshold. This extension has been extended to also provide sentence estimations. In this master’s thesis, these two approaches, that attempt to reduce user effort during the translation session, have been developed. Furthermore, our results using neuronal models have been compared, with those obtained in previous works using statistical models. The results obtained show that there is a greater reduction in the number of words to write when using neural models. |
|---|