University at Buffalo, The State University of New York
Office Address: 338D Davis Hall, Buffalo, NY 14260
Phone: +1 (716) 645-1602
E-mail: rohini@buffalo.edu
There has been much focus on generative language models (LLMs) including Chat GPT and its successors and competitors. LLMs however, are only one component of a complete conversational AI system, or chatbot. Building and deploying chatbots for use with real users in real environments must consider various aspects including:
- Trustworthiness: dealing with issues such as hallucinations, social norms
- Purposeful conversations: chatbots that can encourage critical thinking, debate, counsel, persuade
- Personalized chatbots: chatbots that take into account personal data, chatbots that can speak on behalf of users with disabilities
- Knowledge Grounded chatbots: chatbots that leverage extensive KBs (e.g. UMLS)
- Evaluation methodology
This in turn requires advances on multiple fronts including language understanding and generation, dialogue planning, selecting the best response from an ensemble of response generators, etc. Since conversational data sets are still relatively scarce, it may also be necessary to create data sets for use in developing new models. We will also discuss shifts in model training paradigms including active learning, sample efficient training etc.
This work is partly supported through an NSF grant Purposeful Conversational Agents based on Hierarchical Knowledge Graphs.
Conversational agents (CAs) like GPT-4 have rapidly evolved, showcasing impressive abilities in complex tasks. This progress has spurred debates on AI safety and prompted some researchers to call for cautious exploration. While these agents excel in tasks like coding and passing exams, they lack human-like qualities and struggle with nuanced tasks like persuasion. Traditional persuasion relies on logos, ethos, and pathos appeals, yet most research on persuasive dialogue systems (PDS) focuses on arguments, neglecting human elements. Domain-specific PDS research lacks variety and human-centric interactions, while the absence of well-annotated data impedes progress. This proposal seeks to create a framework for analyzing and generating persuasive conversations, incorporating human characteristics and argument graphs. The plan involves defining features, testing controllable models, constructing argument graphs, and refining response planning in open-domain PDS.
Recent advancements in open-domain conversational agents have harnessed neural response generators and vast, diverse training data, with the Alexa Prize competition showcasing socialbots capable of extensive topic-spanning discussions. Despite their prowess, these systems remain confined to aimless chitchat, beset by artifacts of neural response generation that undermine their reliability and impact, prompting our focus on rectifying open-domain dialog system limitations. The research is structured around three core research questions: (1) ensuring dialogue system consistency encompassing history, emotion, and persona; (2) mitigating hallucinations in knowledge-grounded conversational systems via deep exploration, perturbation strategies, and query rewriting frameworks; (3) enhancing adherence to social norms through data audits and socially aware response generation. The thesis aims to enhance conversational agent coherence, reliability, and prosocial attributes to bolster user trust and enable meaningful societal applications.
We are developing an AI-powered chatbot to revolutionize Augmentative and Alternative Communication (AAC) for patients with Amyotrophic Lateral Sclerosis (ALS). Leveraging the capabilities of state-of-the-art conversational AI models like BlenderBot and DialoGPT, we are finetuning them to generate multiple, contextually-relevant responses to user prompts, thereby offering a more personalized and dynamic communication experience. While the current iteration is text-based, we plan to evolve this system into a multimodal platform, incorporating additional communication mechanisms such as gesture recognition and eye-tracking. Our ultimate aim is to significantly enhance AAC users' communicative autonomy and overall life quality.
The Deception Awareness and Resilience Training (DART) platform helps seniors recognize threats and protect themselves. A collaboration between researchers, game designers, and community organizations, DART is unique in tailoring its curriculum and using gamification to make training accessible and engaging for seniors. The team is currently developing multiple chatbots for the DART project. The first is a chatbot which could help users by answering their queries about the various modules within DART Learn. There is another chatbot which would help users reflect on their learning and experience with disinformation and misinformation. Both the chatbots are deterministic and developed using RASA framework, with plans to add a neural bot and speech recognition soon.
How can we make sure conversational chatbots are faithful to the knowledge and set of facts that we define? In this project we try to tackle the problem of hallucination in large language models with a focus on conversational.