Uso de metáforas por aprendizes brasileiros bilíngues de inglês

Abstract

Research in Learner Corpus Linguistics has traditionally focused mostly on grammatical (e.g., Rankin, 2015) or lexical (e.g., Cobb and Horst, 2015) issues. Consequently, much less is known about how students construct discourse in school tasks. This study aimed to address this gap by investigating the discourses present in written texts by Middle School and First-year High School students in Brazil. To achieve this goal, Lexical Multidimensional Analysis was employed (Berber Sardinha and Fitzsimmons Doolan, 2024) to identify the main discourses in a corpus of 450 texts, amounting to 63,297 words. Lexical Multidimensional Analysis is an off-shoot of Multidimensional Analysis (Biber, 1988), which allows for the detection of the underlying discourses in corpora through the application of multivariate statistical procedures to lexical data extracted from the corpus. However, in the Lexical Multidimensional Analysis literature, little is known about whether metaphorical language is manifested in the dimensions, that is, the extent to which metaphorical language contributes to the dimensionality of texts. Thus, this study aimed to fill this second gap, seeking to determine if the lexical items forming the dimensions are metaphorically used or not. Moreover, in the corpus-based metaphor literature, there has long been a desire to automate metaphor detection, so as to enable the analysis of larger corpora in metaphor studies,, as metaphor analysis in corpora is generally done manually, restricting the amount of data to be analyzed. Given this limitation and the recent availability of Artificial Intelligence through chatbots like ChatGPT, Gemini, and Llama, this study aimed to contribute to text the metaphor detection capabilities of ChatGPT. Eight lexical discursive dimensions were identified through Lexical Multidimensional Analysis. For example, in Dimension 1 "Abstract, theoretical, and scientific knowledge versus Family dynamics, personal resilience, and emotional growth", the discourse in the positive pole emphasizes structured research and scientific analysis, common in academic contexts. On the other hand, the negative pole focused on personal narratives, family relationships, and overcoming personal problems. For the manual annotation of metaphors, the MIP (Pragglejaz Group, 2007) was used, and metaphor candidates were extracted from the variable list of each dimension. At the end of the manual annotation, 245 occurrences (tokens) of candidates were annotated as metaphorical, corresponding to 37% of the candidates (types). Regarding the automatic annotation of metaphors via ChatGPT4, at best one in every four (25%) of the metaphors identified by the human analysts using the MIP protocol (Pragglejaz Group, 2007) was identified automatically through AI. In conclusion, this study contributed to understanding not only the discourses mobilized by students in doing the tasks, but also to how tasks influence the use of linguistic metaphors, as well as to describing the relationship between students' age or proficiency and the incidence of metaphors. Regarding the automation of metaphor identification, the research tried out a range of prompts, charting the process of prompt development, such as restricting the context in which linguistic metaphor occurs. However, the results suggest current LLMs are unable to detect the majority of the metaphors in the corpus