Analysis of context-aware translation memories : part-of-speech pattern distribution and gender neutral translation memories

• Autor(a) principal: Paulo, Marjolene Haidé Machado
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10451/56103
• topic: Domínio/Área Científica::Humanidades::Línguas e Literaturas

Translation Memory is the most commonly used Computer-aided Translation system, whose main purpose is to store and retrieve previous high-quality translated sentences. Although they are very important systems for the translation process, they operate at the sentence level. This approach can be problematic as translated documents without considering the full context can cause coherence and cohesion issues at the text level. The present thesis aims at the analysis and creation of context-independent and gender neutral Translation Memories by resorting to part-of-speech (POS) information, in order to automatically identify context-dependent segments. This would enable the reuse of segments without causing meaning constraints at the document level. In order to achieve this purpose, three experiments were conducted focusing on customer support data. We firstly conducted a pilot experiment to annotate context-related issues in a dataset of 2,045 TMs for Brazilian Portuguese and European Spanish. The results showed that gender agreement was the most frequent category (80%), followed by Register (20%). For the second experiment, we analyzed a total of 5,200 segments for Portuguese, Brazilian Portuguese, Spanish and Latin-american Spanish, with English as the source language. The goal of this analysis was to annotate context related issues with a new context annotation typology. Thereafter, all the context-dependent data was analyzed by a POS tagger in order to understand if it was possible to create sequences of parts of speech patterns that could distinguish context-dependent TMs from context-independent ones. The last experiment consisted on the analysis of a new dataset of a total of 8,000 segments for the same languages. The goal of this experiment was to verify if the previously found patterns could actually identify context-dependent segments. Results showed that 1,298 out of 15,245 TMs were context-dependent in which 1,263 had gender constraints. We were able to turn the latter segments into gender neutral, therefore improving 8% of very frequent data.