The problem of Anglocentrism in the development and training of large generative language models has only recently begun to be considered as an area of study within computer science, specifically within the discipline of natural language processing (NLP), as one of the focal points in the reproduction, propagation, and perpetuation of stereotypes and cultural biases originating in Anglophone culture.
This research reviews the most relevant emerging literature produced by NLP experts, specifically the work of Zhou et al. in their June 2025 paper Culture is Not Trivia, and offers an interdisciplinary perspective from the humanities, specifically from Latin American cultural theory, in order to begin what could be called a practical articulation of their findings. Such an articulation would make possible the creation of language models that are not only technically expert in a culture, but culturally situated: language models trained on localized data and fine-tuned from the center of a culture rather than as spectators of it.
As part of this task, the project will outline a specific theoretical framework for the Latin American region. This framework will provide a solid platform for understanding the specific particularities of our culture, not only in historical terms, but also in order to avoid reducing those characteristics to mere "oddities" of our different dialectal varieties of Spanish, and to understand them instead as their living pulse.
To this end, the research proposes a deeper examination of the categories of "noise" and cultural proxy mentioned in Zhou et al.'s work, as an indispensable step that would make it effectively possible to reach the goal of creating culturally competent systems. It also proposes a methodological review of how both categories are treated within natural language processing pipelines, arguing that in certain contexts this material does not constitute mere residue, but instead carries situated cultural logic. In some cases, this requires a partial resignification of fragments of noise as epistemic proxy, or the creation of an intermediate category that would allow their integration without loss of meaning.
Finally, in resonance and coherence with the challenge this field faces, the present research was originally conceived and written in its target language: Latin American Spanish. Its subsequent translation into English will constitute a gesture that stages the final challenge NLP faces in matters of cultural competence when developing culturally attentive language models.