Conversational AI has come a long way from the basic chatbots that provided scripted responses, only to cause inconvenience and later connect you with a live agent. Today, the technology has evolved tremendously, owing to the advent of neural networks, the transformer model, and OpenAI’s GPT-4. These complex systems, which have found significant usage across several industries, including education, healthcare, finance, and voice assistant technologies, among others, are becoming increasingly helpful in everyday life.
However, the development of these technologies is often driven by technological concerns, and it tends to overlook users’ needs and socio-cultural contexts. According to the Association for Computational Linguistics, or ACL (a scientific and professional organization for people working on natural language processing), this approach, combined with the scarcity of human rights regulation of AI, raises concerns about linguistic discrimination, exclusion, surveillance, and security risks. Additionally, the data used for training conversational AI mostly comes from written rather than interaction-based language data sets and often does not include gestural, social, and emotional aspects that are fundamental to moral human interaction. These considerations raise a need to promote a positive impact of conversational technology on linguistic diversity and inclusion, making it imperative to balance technological concerns with socially relevant matters.
Recent research projects concerning aligning language models with human values are attempting to resolve the challenge of integrating morality in Artificial Intelligence in many ways, including Stanford’s students’ work on Moral Integrity Corpus — a benchmark for ethical dialogue systems, and the University of Edinburgh’s modeling moral and ethical judgment of real-life anecdotes from Reddit, among others.
“But the work remains in its early stages, facing two noteworthy challenges,” says Owen Rambow, a professor in the Department of Linguistics with a joint appointment in IACS (Institute for Advanced Computational Science) at Stony Brook University. “ First, while these models can adapt to new situations, they often struggle to uncover new norms and model richer contexts. Second, the norms covered in these studies are primarily Reddit-based and US-centric.”
To overcome these obstacles, Professor Rambow collaborated with researchers from the University of Illinois Urbana-Champaign and Columbia University to work on NORMSAGE — a novel framework that can automatically extract culture-specific norms from multi-lingual conversations on the fly, as well as provide explainable self-verification to ensure the correctness and relevance of its results. The team tackled the issue of US-centricity by training NORMSAGE on dialogues from TV shows of single-culture (Big Bang Theory), cross-culture (Citizen Khan), and multi-lingual (Real-World Negotiations) origins, as well as multilingual (Chinese and English) conversations from real-world chats, negotiations, and documentaries. The paper, which was presented at EMNLP — a leading conference in the fields of Artificial Intelligence and Natural Language Processing, organized by ACL, also discussed the promise of this approach in extracting high-quality culture-aware norms across several quality metrics, charting the path for more inclusive AI systems.
Related work in overcoming these challenges also includes XDailyDialog, a multilingual parallel dialogue corpus developed by researchers from ETH Zurich, which provides a multilingual, cross-cultural open-domain dataset of dialogues, as well as a generation model, kNN-Chat, with a novel mechanism to support unified response retrieval for monolingual, multilingual, and cross-lingual dialogue.
These projects, among several others, including Luxembourg Institute of Technology’s researchers’ work on socio-cultural adapted chatbots, Alana AI’s research on integrating language variations into conversational AI agents, and University of Antwerp’s students’ discoveries in ‘Conversational AI: Multilingual Capabilities, World Knowledge, and Evaluation Strategies,’ have the potential to draw a broad impact, as they work toward empowering the discovery of norms across cultures, expanding their multilingual capabilities, and enabling individuals to navigate communication barriers, find common grounds, and collaborate effectively.
Communications Assistant
Ankita Nagpal