2025 Theses Doctoral

Structured representations as a lens for meaning: reflecting and emitting content with Abstract Meaning Representation

Lee, Fei-Tzin

When we communicate with one another in language, among the conditions that make it possible for information to be reliably exchanged are two commonalities between the participants. First, we have access to a partial shared knowledge base, with some grounding that makes it possible for both participants to identify what is being referred to. Second, we have a common understanding of the structure of language itself, and more specifically the way in which information held in the mind is compressed into and extracted from language. Neither of these are immediately available to machines. Large pretrained language models appear to learn at least a functional approximation of both of these things. However, even as language models have grown more capable at generating fluent text, accessing world knowledge, and generally performing tasks that in humans we correlate with reasoning, their semantic unreliability has remained a constant. The persistent hallucinations, both pragmatic and factual, produced by language models starkly indicate that neither their understanding of the world nor their understanding of the rules of language at a semantic level are consistent with those of a human (if, indeed, such a thing as ``understanding'' even exists in the context of a language model).

The central question of this dissertation is this: how can we ensure that machines express meaning through language the same way humans do? We approach this problem by introducing an intermediate representation that describes the semantic structures used in the human mind, which we repurpose as an interface to model behavior. When used to analyze and evaluate text, this representation functions as a high-level content sketch that allows us to compare and identify common content across related texts. When used in generation, it becomes a content plan that lays out the desired content of output in a structured format, allowing us to specify surface-level details and modify the plan itself prior to realization.

In this work, we use as our interface Abstract Meaning Representation (AMR), a graph-based semantic representation that captures the types and relationships of concepts described in text. We explore three aspects of the usage of AMR in this role.

First, we study the practicalities of constructing and interpreting AMR graphs: we conduct a detailed analysis of error cases in prior methods of constructing a document-level graph from sentence-level components and design a new method that addresses these issues, yielding improved performance in the content selection phase of summarization.

Second, we explore their utility in directly comparing meaning, rather than form: we develop models for the novel task of AMR node alignment and propose two novel methods of AMR-based text evaluation that in part draw upon these alignment models. Finally, we apply AMR to generation: we demonstrate that using AMR as a content plan, in conjunction with pretrained language models, allows us to produce better-specified and more controllable outputs, at a finer granularity than previous work in controllable generation.