Title: Cultural Differences and Changes Among Native Americans and Their Importance for NLP

Abstract: In this talk, we present the cultural differences and changes among indigenous peoples compared to Western cultures. We will begin with a broad overview of key indigenous cultural aspects before focusing on specific case studies: the Kickapoo, Cahuilla, Wixárika, and Rarámuri communities. Here, we analyze some aspects of the political-religious structure (relationship between spirituality and politics), the social-ceremonial structure, and the socioeconomic activity of these peoples, as well as their changes over time. Ethical considerations are also of great importance when it comes to research on indigenous peoples. This also applies to research done for NLP. The traditions and beliefs of the different tribes and communities must be taken into account. The key is not to consider indigenous peoples as objects of research but rather to create a balance and integrate them into projects. Therefore, it is important that the researcher understands the culture of the indigenous communities to achieve objective results.

Title: Culturally Aware Machines: Why and when are they useful?

Abstract: Language, culture, and technology are deeply intertwined; any technology that aims to process or generate language must have a nuanced understanding of culture. Yet, culture is elusive, often defying concrete definition, which poses a profound challenge to NLP and AI research. While the NLP community has acknowledged this fact and taken initial steps by developing culture-specific benchmarks and identifying biases in large language models (LLMs), there remains a gap in addressing why cultural understanding matters for machines, and the real-world impact on users of culturally (un)aware machines. In this talk, I will ask and attempt to answer the following three questions: (1) What can we do better with AI systems that are culturally aware? (2) Why is modeling culture so complex, and what can we learn from psychology and anthropology? (3) How can LLMs, in turn, be leveraged as tools for studying and enhancing our understanding of cultures?

Title: Engagement and Access in Language Work Across Cultures

Abstract: Language research and development now often aim to proceed in a community-engaged way out of an awareness that local linguistic communities are diverse and disempowered. But what will it take to make the products of language work truly accessible across the world’s full range of settings and cultures? In this presentation I share insights gained through nearly thirty years work with Arapesh people in the Pacific island nation of Papua New Guinea (PNG). PNG is famous for its dramatic linguistic diversity while also being ill-positioned to take advantage of even the most basic linguistic resources and services, and this despite nearly a century of eager missionary and academic linguistic attention. The reasons for this are both infrastructural and cultural, in surprising ways that I will elucidate. Challenges to access are especially unfortunate in PNG, where language shift from local vernaculars like Arapesh to languages of wider communication is now “rampant” (Evans 2010, Kulick and Dobrin in press). Building from the PNG case study, I argue that linguists need an expanded imagination of what community engagement and cultural access might really require.

Title: Cultural value alignment in multilingual and multimodal language models

Abstract: Multilingual language models (MLMs) are trained on the concatenation of text from a wide variety of languages spoken in the world, and thus we can expect different, and perhaps opposing, cultural values to be encoded in them simultaneously. This necessitates MLMs to become inherently multicultural as well in order to appropriately serve culturally diverse communities. In this talk, I will discuss our recent research on cultural value alignment in MLMs, addressing this topic from two perspectives: (1) investigating cultural value shifts during MLM fine-tuning in different languages; and (2) steering cultural value alignment via in-context learning at inference time. In the final study, I will shift the focus to vision-language models, investigating the role of visual information in encoding cultural cues relevant to alignment. I will then discuss the implications of our findings for the design of LLM alignment methods.

Title: Measuring cultural biases in LLM generations

Abstract: TBD