How Does AI Map Meaning Across Languages and Time?

This summer, we invite you to explore the experiences and accomplishments of our 2025-2026 Bass Connections project teams. This series features team-written profiles that showcase the discoveries, challenges and impact of teams who spent the year tackling real-world problems.

Have you ever wondered how meaning is represented, organized and transformed within AI systems? It turns out the answer looks less like language as we know it and more like a shifting map of relationships between concepts.

Before AI can process language, it converts words into numerical representations called “embeddings,” where words with similar meanings are positioned closer together. The Historical Shifts and Geographical Drifts: An Exploration of AI Embedding Space project team explored how these meanings – and their respective positions as they are represented in AI – shift across time and languages. The goal was to create accessible ways to visualize those changes. 

Subteams focused on shifts in meaning over the past 30 years, differences in how concepts are represented across languages and creative ways to make these patterns visible to the public. The team’s work highlighted key questions about how AI represents complex concepts mathematically, why these shifts matter and how they can be communicated in accessible and engaging ways. To share their findings, the team developed an interactive exhibit that invited audiences to experience and explore embedding space firsthand.

This team was led by Brinnae Bent (Pratt School of Engineering) and William Seaman (Arts & Sciences: Art, Art History and Visual Studies).

Image

By: Members of the Historical Shifts and Geographical Drifts: An Exploration of AI Embedding team

This project brought together an interdisciplinary team spanning computer science, engineering, media and visual arts, statistics, linguistics, design and the social sciences. We worked in three subteams, each focusing on a different aspect of the project. The temporal team examined how AI-related and value-laden words evolve over time, tracing how their meanings have shifted over the past 30 years across different contexts. The geolingual team asked whether AI systems represent the same concepts consistently across languages, or whether meanings drift depending on linguistic and cultural context. The arts team brought this research together in an interactive exhibit designed to make complex ideas something people could walk through, explore, and experience.

The goal of the project was to better understand how AI systems create and use meaning, and to find ways to measure, explain and display that process clearly. We focused on how meaning changes over time and across languages, working to make those shifts visible and easy to grasp. Central to this effort was the idea of “embedding space,” where concepts are represented as points in a map-like structure, with similar ideas placed close together and different ones farther apart. The arts team translated this abstract system into an interactive, accessible experience for a broad audience.

Image

The project culminated in a two-part interactive exhibit, “Lost in Embedding Space,” at the Rubenstein Arts Center. One installation visualized the temporal team’s work as a walkable timeline from 1995 to 2024, allowing users to see how AI-related language has evolved. The second translated the geolingual analysis into tree structures, where each concept appeared as a set of branches representing different languages, showing how meanings diverge or align across cultures. The installations combined sensors, 3D printing, and visual projections to create an immersive experience.

Alongside the exhibit, the team developed pipelines to support both lines of research. The temporal team built a visualization pipeline by fine-tuning embedding models on 30 years of academic and public data to track changes in word meaning over time. The geolingual team constructed a multi-stage pipeline to measure how concepts shift across languages. They built a concept inventory, collected parallel sentences in six languages and embedded them with multiple multilingual AI models. These analyses produced both quantitative results and visualization-ready outputs that directly informed the exhibit.

The team also created a video to capture the story behind the project. In it, team members reflected on the research, the collaboration, the excitement of interdisciplinary work and the process of turning abstract machine learning ideas into a public-facing experience.

What We Learned

These reflections have been lightly edited for length and clarity.

Qifeng Cheng (Ph.D. in Physics ’29)

This Bass Connections project became, for me, a way of rethinking what it means to understand science, and what it means to share that understanding with others. Working with AI embeddings and data, we began with technical questions, but quickly found ourselves asking something deeper – how can this knowledge be felt, not just explained? Being part of this group made this question real. It made me notice how much of science is often left untranslated when we speak outside our own field. And it was through this amazing teamwork spanning from modeling, training, data work, and visual and physical design that made these gaps visible and bridgeable. The way we supported and stepped in for one another made me feel stronger and grateful. I deeply appreciate the opportunity to be part of this work. It has motivated me to continue asking and answering questions at the interaction of science and art, and about how knowledge can be shared in ways that are both meaningful and human. 

Image

Bochu Ding (Master of Engineering ’26)

What an inspiring experience! For me, Bass Connections was less about “what” and more about “who.”  I was always in awe of what my teammates brought from their corners of expertise: from Maddie’s prowess with TouchDesigner and Alexis’ concepts and sketches, to James’ knowledge of abstract math concepts and Qifeng’s knack for physical prototyping. There’s no other team I would have trusted to blend technical concepts, cultural sensitivity and creative construction in such a formidable way. And of course, I can’t thank our instructors enough, especially Dr. Bent who encouraged blue-sky exploration and who spent many hours untangling ideas with me. 

Junyu Zhang (Master of Engineering ’26)

Being part of this Bass Connections team was one of the most meaningful experiences I have had at Duke. As an international student and an introvert, I really appreciated how welcoming and inclusive the whole team was. My teammates, Professor Brinnae Bent and Professor Bill Seaman created an environment where I felt comfortable contributing ideas and learning from people with very different backgrounds. As a multilingual person, I was especially excited by the way this project connected technical AI methods, such as embeddings and large language models, with broader questions about language, culture and fairness. The project helped me see that AI is not only a technical system, but also something deeply connected to society, identity and communication across cultures. This experience gave me a clearer direction for my future Ph.D. research interests, especially regarding how AI can become fairer and how AI research can connect with social science, business and real-world human behavior.

Alexis Golart (Statistics ’28)

Being a part of Bass Connections was an incredible opportunity to engage in research and gain exposure to an area I likely wouldn’t have explored so deeply otherwise. Working on the arts team allowed me to develop a stronger understanding of transformer models and embedding spaces, while approaching the material through a creative lens that I was originally passionate about. I especially appreciated the collaborative nature of the team and how it brought together undergraduates, master’s students and Ph.D. candidates who had such a wide range of experiences. The non-arts teams also played a huge role in supporting our work leading up to the exhibit, which made the experience really special. Bass felt like one of the closest reflections of real-world experience and collaboration that I’ve had during my time at Duke University.

Neha Shukla (Computer Science ’27)

Image

I’ve absolutely loved being a part of this Bass Connections team! This has been such an interdisciplinary learning experience with the most close-knit team, and I’m incredibly grateful for the mentorship and guidance from Dr. Bent and Dr. Seaman every step of the way. I’ve grown so much as a quantitative researcher and was so excited to connect our technical findings to broader questions about algorithmic fairness and the societal impacts of systemic bias. It was so much fun bringing the ideas we sketched and prototyped to life, and I loved translating the research I was conducting to building 3D models, circuits and sensors, interpretable visuals and an interactive art exhibit. This journey has taught me to find resilience as a researcher and deepened my love for sticking with a problem and chipping away at finding solutions. Seeing our interactive exhibit come together and the public engaging with our findings felt like the most special culmination of all of our progress over the year. I’m so grateful to our incredible professors for giving us true freedom to explore and create, and to the best team for bringing our vision to life! 

James Sohigian (Mathematics and Computer Science ’28)

This Bass Connections project bridged my theoretical understanding of word embeddings that I learned in the classroom with interpretable and practical applications of real-world data. Over the course of the year, I gained insight into model fine-tuning, embedding spaces, dataset cleaning, dimension reduction techniques and transformer architectures. The technical lessons I learned, combined with the creative and artistic thinking that my teammates exposed me to, will be with me forever. I truly appreciate the uniqueness of the project, the amazing professors who guided us and my amazing teammates!

Maddie Tsang (Computer Science and Visual Media Studies ’28)

Bass Connections was an amazing opportunity to explore interdisciplinary research and learning in such a close-knit setting. Our project was so special in how we took data and transformed it with creativity, turning embedding into emotions. I am so grateful for all of my team members and professors who were endlessly supportive and inspiring. It was amazing to see how everyone’s unique expertise came together to create our final installation. I had so much fun collaborating and creating with this team and given the opportunity I would love to do it again!

Learn More

• Browse additional news and updates from Bass Connections.
• Explore additional teams in the Information, Society & Culture theme.
• Learn about the project team experience through stories from students.

Main image: Members of the public view the Geolingual Growth Art Installation.