Abstract
Underspecification is a general property of language: the surface of a text rarely fixes thefull state of the world it describes. A system can fill that gap with a statistical prior, the
most frequent completion in its training distribution, or resolve it from structure, computing
the answer the particular expression licenses rather than retrieving the frequent one. The
first is the statistical understanding that current language models embody and that static
benchmarks reward; the second is the compositional understanding long pursued in symbolic
AI and computational semantics. This dissertation pursues a neuro-symbolic synthesis of the
two: it builds an explicit event–entity structure and lets operators of more than one kind,
symbolic and data-driven, run over it, so that structure supplies what can be inspected and
audited while the data-driven machinery supplies what hand-written rules cannot, a graded
notion of type compatibility in place of an exact match or none, and coverage learned from
data rather than added one rule at a time. Procedural texts such as cooking recipes are
the proving ground, because they make the phenomena at stake dense and interdependent
on the surface: results left implicit, one entity recurring across steps under transformation-
induced renaming, and state carried along a chain of actions. The representation is the
Process-Oriented Event Model (POEM), a Generative Lexicon account of procedural events,
derived from existing expert-built lexical resources rather than from task-specific supervision
or unconstrained language-model generation, and independent of any one implementation of
its operators. The work has three parts.
Chapter 2 establishes the empirical and methodological ground through two threads
of annotation and data work central to this dissertation. The first is a series of manual
procedural-annotation corpora, built on the R2VQ recipe corpus and extended through
CUTL, where the Process-Oriented Event Model was first introduced, and GLAMR, which
ports it onto mainstream meaning representation. The second is OpenPI and its language-
model-augmented extension OpenPI-A, produced through an inverse-annotation method that
recovers event-trigger annotations from existing state-change annotations and extends the
framework beyond cooking to broader procedural genres. Together these resources establish a
theoretically grounded annotation target and a methodological pivot toward deriving POEM-
style structure without relying on either expert annotation or unconstrained language-model
generation alone.
Chapter 3 builds a symbolic model of event semantics, framing its composition as dis-
tributed compositionality: a predicate and its arguments each supply part of the meaning,
read from existing lexical resources and the codified human expert knowledge they encode
without any task-specific annotation. Two complementary streams feed the model: an
event-side stream reading each predicate from VerbNet-GL, and an entity-side stream reading
each argument through a cascade of lexical resources (BSO, CoreLex, WordNet, and BECL).
A compositional operator, built from the outer and inner application of Generative Lexicon
and given here a symbolic, rule-based implementation, then combines the two distributed
contributions to construct, for each event, the Generalized Result Role (GRR): the symbolic
representation of an event’s output, whether textually realized or implicit. Chained across
series of events, these result roles form the structural backbone of the representation and
anchor the operationalization of the GL-based Dynamic Object Model without any manual
event-level labeling.
Chapter 4 specifies a second, vector-symbolic realization of the same operators, for the
long tail the symbolic rules cannot reach, where an exact type match gives way to graded
compatibility in a vector space. Recast as typed records, POEM’s constructs are enumerable,
so the finite lexicon generates its own gold-labelled corpus, which serves both as a probe
put to instruction-tuned models and as the operators’ eventual supervision. Each event’s
result is reified into an atomic entity a later event binds directly, and the operators become
a vector-symbolic architecture whose role vectors are the learnable part. This architecture
is specified and justified as a design, its implementation left as the immediate future work;
only the symbolic pipeline of Chapter 3 is measured on CUTL, which remains the design’s
eventual target.
In sum, this dissertation advances procedural-text understanding by treating a demanding,
theoretically grounded representation (one that unifies implicit argument detection, coreference
under transformation, and entity state tracking) as a single compositional target.
By
combining the empirical grounding of prior annotation work, a symbolic pipeline derived
from expert lexical resources, and a vector-symbolic design that makes the same structure
learnable while keeping it inspectable, it shows how a knowledge-rich symbolic pipeline and a
specified data-driven architecture together frame rich, evaluable event–entity structures from
natural-language procedures, and points toward grounding procedural text in embodied and
multimodal settings, tasks that reward the explicit, accountable world state such a structure
makes available.