Artificial intelligence (AI) systems, especially those powered by large language models (LLMs), are currently experiencing prosperity and rapid development. However, LLMs still show reliability deficiencies of understanding and applying knowledge in various domains or contexts. This indicates room for improvement in LLMs' capabilities of reasoning, which typically requires implicit thinking and integration of world knowledge to infer dynamic output actions or results, rather than direct recourse to memorized information. This thesis aims to model such implicit inferences as structured schemas, and improve the reasoning reliability of LLMs by incorporating the modeled inference schemas as intermediate scaffolds. Our studied reasoning tasks cover both open-ended narratives such as dialogue modeling and deterministic question answering (QA) such as mathematical problem solving.
We first build data resources to benchmark non-standardized inference schemas underlying open-ended narratives. Specifically, we construct a persona-grounded commonsense knowledge graph (KG), PeaCoK, to collect commonsense facts of persona interactions that ground human narratives. We then build a contextual knowledge linking dataset, ComFact, to benchmark contextually-relevant commonsense facts for narrative understanding and generation, which form the required inference schemas in narratives. (Part I)
We further develop methods for AI systems to learn the inference schemas in narrative contexts. First, we train neural knowledge linkers based on our ComFact dataset, denoted as ComFact-Linker, to retrieve the commonsense facts (e.g., the facts in our PeaCoK KG) relevant to the narrative contexts, which augment the consistency and engagement of downstream dialogue generation. Noticing the inefficiency of knowledge retrieval, we further develop a family of diffusion-based knowledge models, DiffuCOMET, which can directly generate commonsense facts with better contextual relevance and novelty, compared to autoregressive knowledge models. (Part II)
As an extension to more data modalities, we also propose a visual narrative benchmark, VinaBench, which annotates a set of commonsense and discourse knowledge constraints as inference schemas in textual-to-visual narrative generation. Based on that, we then train a visual narrative constraint generator, VinaCoG, whose output knowledge constraints effectively improve the visual narrative consistency and alignment to the input textual narrative. (Part III)
In terms of deterministic QA, we first propose a chain-of-abstraction (CoA) method, where LLMs are trained to construct abstract reasoning chains as general inference schemas. CoA enables LLM agents to more efficiently use external tools, and meanwhile improves their reasoning accuracy, in both mathematical reasoning and Wikipedia-based QA. We also develop an abstraction learning framework, AbstRaL, which promotes LLMs to construct more faithful abstract reasoning chains, by using g
