Adversarial strategies for Reducing Gender Bias in Neural Machine Translation
In a more connected world, communication between different native speakers has became more necessary. This make that translation systems become more useful. In the last years, typical translation systems have evolved towards NMT, that have achieved State of the Art results. NMT models use deep learn...
| Autor: | |
|---|---|
| Formato: | tesis de maestría |
| Fecha de publicación: | 2020 |
| País: | España |
| Recursos: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/334929 |
| Acesso em linha: | https://hdl.handle.net/2117/334929 |
| Access Level: | acceso abierto |
| Palavra-chave: | Machine learning Neural networks (Computer science) NLP Deep Learning Bias Machine Trnslation Aprenentatge automàtic Xarxes neuronals (Informàtica) Àrees temàtiques de la UPC::Enginyeria de la telecomunicació |
| Resumo: | In a more connected world, communication between different native speakers has became more necessary. This make that translation systems become more useful. In the last years, typical translation systems have evolved towards NMT, that have achieved State of the Art results. NMT models use deep learning algorithms, that are trained with huge datasets, that contain the social bias and stereotypes (e.g gender, race, religion, age...). So this implies that the model implicitly learns those biases. In this thesis we will see tat social stereotypes are present in NMT and how to mitigate the gender bias that appears. We try to face this problem using an adversarial strategy. We propose a model architecture composed of two blocks, the generator a NMT model and a discriminator that decides whether the translation keep the gender or not. We evaluate our proposed system on the test dataset, and show that translation performance gains on accuracy and F-Score . This means that our proposed system have learned to neutralize previously existing biases on the baseline system. |
|---|