Title: Analysis of semantic aspects in automatic translations of texts
Advisor: Gustavo Paiva Guedes e Silva
Committee: Gustavo Paiva Guedes e Silva (Cefet/RJ) (President), Eduardo Bezerra da Silva (CEFET/RJ), Eduardo Soares Ogasawara (Cefet/RJ) and Lilian Vieira Ferrari (UFRJ)
Day/Time: July 4, 2018/10h
Room: Auditorium V
Automatic text translations emerged in the years 50, motivated by military issues. Nowadays this type of translation is part of our daily life and represents an important tool for communication in the globalized world, especially with the use of tools for automatic translation of texts available in the Web environment. However, despite being an area with more than 60 years of study, there are still several challenges to be overcome, which makes this type of process remain dependent on human review. There are currently several metrics to evaluate automatic translations of texts, among which, the BLEU metric is presented as state of the art. This metric evaluates the quality of the translations based on the exact and orderly matching of words, without considering. However, the semantics (e.g., linguistic and psychological aspects) of the sentences evaluated. In this scenario, the main objective of this work is to propose a new metric that can add semantics to the assessments of this type of translation. As a secondary objective, this work also contributes with two algorithms for aid in the identification and quantification of psycholinguistic aspects in English translations to
Brazilian Portuguese. To achieve the proposed objectives, this work uses an affective lexicon present in a tool called LIWC (Linguistic Inquiry and Word Count). This lexicon is capable of accounting for words in categories that represent psychological aspects and
Language. During the experiments, ten texts translated by two human specialists and three of the tools above were used. The texts above were used to establish a comparison between the proposed metrics and state of the art. The tests also aimed to evaluate possible problems produced by the tools used to perform this type of translation. The results were considered promising and indicated that this study could contribute to new work directed to the development of metrics for the evaluation of automatic translations of texts and perhaps even for works directed to the Development of tools that produce this type of translation.