Learning multilingual and multimodal representations with language-specific encoders and decoders for machine translation
This thesis aims to study different language-specific approaches for Multilingual Machine Translation without parameter sharing and their properties compared to the current state-of-the-art based on parameter-sharing. We define Multilingual Machine Translation as the task that focuses on methods to...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2022 |
| País: | España |
| Institución: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/673982 |
| Acceso en línea: | http://hdl.handle.net/10803/673982 https://dx.doi.org/10.5821/dissertation-2117-365523 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Enginyeria de la telecomunicació 004 621.3 81 |
| Sumario: | This thesis aims to study different language-specific approaches for Multilingual Machine Translation without parameter sharing and their properties compared to the current state-of-the-art based on parameter-sharing. We define Multilingual Machine Translation as the task that focuses on methods to translate between several pairs of languages in a single system. It has been widely studied in recent years due to its ability to easily scale to more languages, even between pairs never seen together during training (zero-shot translation). Several architectures have been proposed to tackle this problem with varying amounts of shared parameters between languages. Current state-of-the-art systems focus on a single sequence-to-sequence architecture where all languages share the complete set of parameters, including the token representation. While this has proven convenient for transfer learning, it makes it challenging to incorporate new languages into the trained model as all languages depend on the same parameters. What all proposed architectures have in common is enforcing a shared presentation space between languages. Specifically, during this work, we will employ as representation the final output of the encoders that the decoders will use to perform cross-attention. Having a shared space reduces noise as similar sentences at semantic level produce similar vectorial representations, helping the decoders process representations from several languages. This semantic representation is particularly important for zero-shot translation as the representation similarity to the languages pairs seen during training is key to reducing ambiguity between languages and obtaining good translation performance. This thesis is structured in three main blocks, focused on different scenarios of this task. Firstly, we propose a training method that enforces a common representation for bilingual training and a procedure to extend it to new languages efficiently. Secondly, we propose another training method that allows this representation to be learned directly on multilingual data and can be equally extended to new languages. Thirdly, we show that the proposed multilingual architecture is not limited only to textual languages. We extend our method to new data modalities by adding speech encoders, performing Spoken Language Translation, including Zero-Shot, to all the supported languages. Our main results show that the common intermediate representation is achievable in this scenario, matching the performance of previously shared systems while allowing the addition of new languages or data modalities efficiently without negative transfer learning to the previous languages or retraining the system. |
|---|