CAREER: Explanation-based Optimization of Diversified Information Retrieval to Enhance AI Systems

University Of Massachusetts Amherst, Amherst MA

Investigators

Abstract

Large generative artificial intelligence (AI) models, such as ChatGPT, are widely used for information seeking—helping people find information on a topic. Compared to traditional search engines, they provide a coherent narrative, which could potentially facilitate the exploratory phase of users' searches. Generative AI responses are more readable, coherent, and contextually appropriate; hence, they sound authoritative and definitive. However, existing generative AI models are subject to problems such as hallucinations, unsupported misleading answers, outright misinformation, and hidden biases. Another issue is that the majority of user queries are ambiguous. Current systems, including those that employ generative AI models, do not appropriately consider ambiguity by providing users with alternative answers to their queries. The vision of this project is to enable users to use generative AI models to obtain an interpretable, diverse, and unbiased set of alternative answers, viewpoints, subtopics, or aspects as required for various questions or tasks in information access, where each distinct answer or viewpoint is faithfully attributable to a set of evidence and supporting information sources. This project aims to make information access easier, more effective, and more trustworthy for users. Given that search is among the most common online activities, this project is positioned to have a substantial impact on society, promoting a more comprehensive understanding of topics, encouraging critical thinking, and facilitating informed decision-making. To achieve the above goal, this project proposes the development of novel retrieval models to enhance the relevance, diversity, and interpretability of their results. This project will develop models for multi-granular diversification of search results to significantly improve the generalizability of retrieval models in providing diverse results for open-domain queries. In addition, this project enables the full utilization of search results by AI systems through explanations of their relevance and diversity. Building on top of explainable search results, the project introduces explanation-based optimization of search results. This involves improving search results based on reasoning over failures of retrieval models. The resulting retrieval systems will be particularly useful for augmenting large generative AI models through access to explainable explicit knowledge. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.