Human-centric trustworthy foundation model reasoning

Fung, Yi

Permalink

https://hdl.handle.net/2142/127440 Copy

Description

Title

Human-centric trustworthy foundation model reasoning

Author(s)

Fung, Yi

Issue Date

2024-09-30

Director of Research (if dissertation) or Advisor (if thesis)

Ji, Heng

Doctoral Committee Chair(s)

Ji, Heng

Committee Member(s)

• Zhai, Chengxiang
• Hakkani-Tur, Dilek
• McKeown, Kathleen
• Hovy, Eduard

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Human-Centric Trustworthy NLP
• Multimedia Knowledge Reasoning

Abstract

In our current era of globalization and digital expansion, we face an overwhelming amount of new data each day, which causes difficulties for us to understand and convey information effectively. Advancements in AI have given rise to language models (LMs) being increasingly adopted in assisting information understanding and communication for different task settings. However, the potential that LMs can serve in supporting human communication is still hindered by important issues. As common examples, LMs may generate vague, irrelevant, or illogical responses, which limits their helpfulness; they may fail to identify misinformation or spread non-factual hallucinated content, compromising information honesty; and they may produce harmful content that unintentionally reflects biases or insensitivities. In response to these pressing concerns, this dissertation aims to address the alignment of language models and multimedia reasoning frameworks with human needs and values. In particular, we explore the research questions of what core task objectives should be defined for LM models to achieve in developing human-centric trustworthy AI, and what model reasoning paradigm better push forward this motivation. Central to our dissertation is the goal of developing novel socially-situated human-model aligned natural language and multimedia knowledge reasoning framework with grounded mechanisms and modularized operations for ensuring not only helpfulness (i.e., provide users with contextually relevant and precise information that meets their specific needs), but also honesty (i.e., prevent the spread of misinformation and build trust between AI systems and users) and harmlessness (i.e., avoid unintended bias, cultural insensitivity, and potential harm). Fundamentally, we observe that while state-of-the-art large language models (LLMs) are proficient in generating and retrieving general information, they often lack the depth of understanding required to navigate less common or nuanced contexts. This deficiency arises due to their reliance on dominant patterns of frequently encountered knowledge during pretraining, rather than deeper connections of insight. As a result, LLMs that undergo end-to-end training alone commonly struggles with issues such as misinformation, shallow understanding, and lack of adaptability in dynamically changing environments and complex sociocultural contexts. Based on this intuition, we introduce knowledge extraction upfront to shortcut relevant information for knowledge-enhanced model reasoning as a unifying paradigm across helpfulness, honesty, and harmlessness reasoning tasks, through guidance principle of information selection with respect to Grice’s Maxim on cooperative principles, to anchor model knowledge to human context and preference with corresponding weights. In this way, passive data ingestion is shifted to active, structured knowledge acquisition and adaptive contextual understanding, enabling more reliable and contextually aware LLM-assisted information communication. To showcase the universal effectiveness of this knowledge extraction information shortcut paradigm, an underlying key aspect involves identifying which guidance principle makes most sense to utilize across tasks centered around promoting information helpfulness, honesty, and harmlessness. Our dissertation investigates solutions to this from the following angles: * Question Answering and Question Generation through LM Grounded Reasoning and Human Intent Modeling First, we identify the main limitations in model reasoning as due to a lack of structured reasoning to bridge information interconnections and enhance localized understanding, and insufficient awareness of human intent. To address these fundamental limitations, we explore novel mechanisms for grounded reasoning by leveraging semantic interconnections, as well as for predicting latent human intents, to enrich contextual representations and improve model inference in complex data domains. * Misinformation Detection and Control As further study on grounded reasoning's importance and challenges, we investigate extending beyond traditional document-level objective functions to develop fine-grained misinformation detection frameworks that take advantage of extracted information network for crucially enabling greater precision and explainability in identifying inconsistencies and misleading information. * Massively Multi-Lingual Multi-Modal Multi-Cultural Norms Understanding Finally, perception of human intents and actions vary by social and cultural groups. We introduce complementary approaches for multicultural knowledge and norm discovery, from noisy conversation on-the-fly and culture-oriented curated document sources, and leverage these norm-relevant cultural knowledge for improving model reasoning in norm adherence/violation detection as well as debiasing. Through systematically addressing these subproblems, this dissertation aims to establish more effective and robust frameworks for socially-situated human-model aligned natural language and multimedia knowledge reasoning, in contribution towards fostering healthy information communication systems. The broader impact of our research is that it enhances AI-enabled information access and communication with greater factuality, reliability, and trustworthiness, from a technical point of view. From an application perspective, it fosters more meaningful human-AI interactions in contribution to our interconnected world.

Graduation Semester

2024-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/127440

Copyright and License Information

Copyright 2024 Yi Fung

Owning Collections