Uploaded May 2024 | Updated September 2026, 4 hours ago
Abstract: Today, search tools and language models are better than ever at directing users to the relevant information according to their needs. However, it remains difficult for the users to put the information in the context of other sources. The lack of proper infrastructures and tools to provide “contextualized” information access has arguably become the Achilles heel of many challenges we face today in the NLP community, such as hallucination and factuality estimation of generated text, as well as how AI reasoners access and consumes knowledge.
In this talk, we argue that the fundamental challenges lie in the way we index and represent textual information. Motivated by how humans process, consume, and infer knowledge, we rethink the linguistic notion of propositions as an abstraction to provide finer-grained access to knowledge in text. We show that propositions can be used as a semi-parametric form of text representation, to allow for contextualized information access across various applications, such as text attribution, retrieval, hallucination detection, semantic representation of text, and so on. Finally, I will discuss future and ongoing work, to study how a more intelligent and contextualized access to information can help AI agents in their reasoning, planning, and decision-making capabilities.
Bio: Sihao Chen is a fifth-year Ph.D. candidate at the University of Pennsylvania, working in the space of natural language process and machine learning. His research focuses on building language technologies to provide more intelligent, contextualized, and trustworthy information access for both users and AI reasoning and decision-making systems. Previously, Sihao worked as a student researcher for Google, and a research intern for Tencent AI Lab.
Abstract: Today, search tools and language models are better than ever at directing users to the relevant information according to their needs. However, it remains difficult for the users to put the information in the context of other sources. The lack of proper infrastructures and tools to provide “contextualized” information access has arguably become the Achilles heel of many challenges we face today in the NLP community, such as hallucination and factuality estimation of generated text, as well as how AI reasoners access and consumes knowledge.
In this talk, we argue that the fundamental challenges lie in the way we index and represent textual information. Motivated by how humans process, consume, and infer knowledge, we rethink the linguistic notion of propositions as an abstraction to provide finer-grained access to knowledge in text. We show that propositions can be used as a semi-parametric form of text representation, to allow for contextualized information access across various applications, such as text attribution, retrieval, hallucination detection, semantic representation of text, and so on. Finally, I will discuss future and ongoing work, to study how a more intelligent and contextualized access to information can help AI agents in their reasoning, planning, and decision-making capabilities.
Bio: Sihao Chen is a fifth-year Ph.D. candidate at the University of Pennsylvania, working in the space of natural language process and machine learning. His research focuses on building language technologies to provide more intelligent, contextualized, and trustworthy information access for both users and AI reasoning and decision-making systems. Previously, Sihao worked as a student researcher for Google, and a research intern for Tencent AI Lab.
