Analisi dei Gender Bias nella Machine Translation: Il caso delle Job titles e degli annunci di lavoro

This dissertation deals with the issue of gender bias in Neural Machine Translation (NMT). This theme is of particular importance in today's world, where technology is present in all aspects of daily life, ranging from leisure activities to work. It should be noted that, despite bringing multiple advantages in terms of time and effort, MT systems still make several mistakes, thus making professional translators indispensable in order to obtain a perfect output in the target language. Some of the most common errors produced by NMT include the translation of proper nouns, “false friends”, and acronyms, the respect of grammatical word order of the target language, and the inability to detect and translate new terms or choosing the right meaning in the case of polysemy. (Goutte et al., 2009) In the first chapter, I will focus on what Machine Translation systems are, in order to provide foundational knowledge of their functionalities and processes, as well as their developments over the years. The advantages and disadvantages that NMT brings to professional translators will be discussed. Subsequently, the chapter will focus on the errors NMT systems make in the transposition between different languages. I will also explore the role and the importance of post-editing, namely the activity that makes it possible to comprehend texts produced by MT systems. In the second chapter, the social problem of gender bias will be presented with its representations in the language and, of course, in the translations provided by MT systems. In fact, the focus of this dissertation will be on the issues linked to the translation of gender bias, defined by the APA Dictionary of Psychology as: “any one of a variety of stereotypical beliefs or biases about individuals on the basis of their gender. These biases can be expressed linguistically, as in use of the phrase physicians and their wives (instead of physicians and their partners, which avoids the implication that physicians must be male/masculine) or of the use of gender pronouns when people of all genders are being discussed”. (APA Dictionary of Psychology, 2023) Concrete examples will be presented in order to accompany the explanation and make it clearer. This issue is important because language is the means through which we express ourselves, and gender bias shows how some kinds of stereotypes still shape society and our perception of different roles and possibilities for different groups of people. Furthermore, it is important to remember that NMT’s output is based on data provided by human professionals and developers, demonstrating their unconscious biases that will appear in the final product. In the third chapter examples of output from different NMT systems will be discussed in order to show how they translate genders in different contexts. Post-editing will be performed on the same texts in order to propose possible solutions to address the issues identified. In conclusion, the solutions provided by some MT systems will be presented, showing how developers are constantly trying to improve their programs. Through this work, I will be able to demonstrate how translators and post-editors can correct the errors made by automated systems by using their linguistic and cultural knowledge, trying to change, in this way, the stereotyped vision of roles and gender that is still widespread in today's society.