Linguistics4fairness: neutralizing Gender Bias in neural machine translation by introducing linguistic knowledge
Neural Machine Translation has the power of learning from a large collection of data, which allows it to learn translations effectively and without requiring linguistic knowledge from the languages to translate. The main drawback is that this large collection of data may not exist (i.e. low resource...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2021 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/348437 |
| Acceso en línea: | https://hdl.handle.net/2117/348437 |
| Access Level: | acceso abierto |
| Palabra clave: | Machine translating gender bias transformer factored transformer linguistic input machine translation linguistic information factored machine translation Traducció automàtica Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial |
| Sumario: | Neural Machine Translation has the power of learning from a large collection of data, which allows it to learn translations effectively and without requiring linguistic knowledge from the languages to translate. The main drawback is that this large collection of data may not exist (i.e. low resourced languages) or this data reproduces the social biases existing in the society. While introducing linguistic knowledge has been widely studied to reduce the impact of lack of data, it may also be interesting to use it as a potential tool to mitigate and neutralize biases. This project explores the Factored Transformer which is a neural machine translation architecture which allows for introducing linguistic features and its impact in gender bias mitigation. In order to study the gender bias in the results of machine translation models, we use the standard WinoMT framework by Stanovsky et al. (2019) that permits to detect gender bias in the translations from English to languages with grammatical gender. We investigate the influence of adding linguistic factors to Factored Transformer on 4 language pairs: English-French, English- Spanish, English-German and English-Russian and detect improvements using several types of linguistic input. Besides general gender bias metrics, we use the methodology described in Costa-jussà et al. (2020) for the interpretability analysis of contextual source embeddings and encoder-decoder attention. |
|---|