Publication Type

PhD Dissertation

Version

publishedVersion

Publication Date

5-2026

Abstract

In this dissertation, we investigate interpretability in the three elements of learning neural text representations: inputs, passed into models, to produce probabilistic outputs. We emphasise perspectives as we present alternative novel methods to mine and organise meaning in this work.

Models. We initiate our investigation by examining Neural Topic Models (NTM), proposing an alternate angle of interpreting its word-topic distribution, producing better topic representations for interpretation. Our method maps the problem of finding these better interpretations to classical NP-hard graph problems, enabling examination of topic distributions in a composite manner. Next, we apply our previous findings to extract interpretations from the weights of transformer-based Large Language Models (LLMs). The theory of concept superposition in LLMs' neurons suggests an impediment to interpretability. We propose to disentangle these superposed concepts, mapping this problem to a classical NP-hard graph problem, optimising on automated coherence metric scores.

Outputs. Noting how we judge observations is critical to interpretability evaluations, we examine text representations from the human mental model. We propose and formulate a large-scale correlation analysis and accompanying user studies to examine automated coherence metrics and human evaluations. Our results show that automated coherence metrics are correlated with human evaluations using text statistics from Wikipedia-English. Further examinations, anchoring its analysis on word statistics, obtain some insights applicable to future user studies.

Inputs. The model and its corresponding outputs learn from the pre-defined boundaries of the input space. For neural text representations, that would be the token space consisting of words and subwords. To better understand the subword space, we propose an interpretation problem to the word space. Finally, we propose a novel perspective on solving for the token space, choosing which word and subwords to include in the limited token vocabulary space.

Degree Awarded

PhD in Computer Science

Discipline

Artificial Intelligence and Robotics | Computer Sciences

Supervisor(s)

LAUW, Hady Wirawan

First Page

1

Last Page

265

Publisher

Singapore Management University

City or Country

Singapore

Copyright Owner and License

Author

Download

Share

COinS