Impacto da qualidade de textos de partida criados por utilizadores e agentes e a propagação de erros em sistemas de Tradução Automática

This paper proposes a typology concerning errors and linguistic structures found in the source text that have an impact on Machine Translation (MT). The main objectives of this project were firstly, to make comparison between error typologies analyze them according their suitability; annotated data build data-driven while adapting previous existing typologies; distinction produced by users agents online Customer Support domain; test proposed with three case studies; methodize patterns verify MT systems; finally, create ready for production its particular field. At first, it was made different typologies, whether they consider bilingual or monolingual level (e.g. Unbabel Error Typology, MQM Typology (Lommel et al., 2014b) SCATE Taxonomy (Tezcan 2017). allowed us differences similarities and, also, which issue types been previously used. In order typology, both sides analyzed — user agent as present writing are influenced factors. results analysis assessed through annotation process calculated one most highly used manual evaluation metrics translation quality Multidimensional Quality Metrics (MQM), QTLaunchPad (2014), funded European Union. Through analysis, then possible Source Typology. aid future annotators we provided guidelines elaborate new additions typology. interest confirming reliability studies conducted internal pilot, total 26,855 words, 2802 239 (represented ‘Neutral’ severity associated conversational markers, segmentation, emoticons, etc., characteristics oral speech) annotated, purposes taking into account several language pairs. these studies, verified effectiveness additions, well transfer target text. Besides that, also had fact any systems. testing confirm including what needs improvement.