AI-assisted_textuality: Enhancing Literary Analysis through Neural Topic Modeling and Visualization

Thesis Abstract

This study investigates how neural topic modeling and visualization can augment literary analysis by enabling scalable, hypothesis-driven exploration of large corpora while preserving interpretive depth and theoretical nuance. The central problem addressed is the limitation of traditional close reading and manual coding in handling expansive digitized literary corpora, which often leads to selective interpretation and inconsistent comparability across texts, authors, and periods. The aim is to develop and evaluate an AI-assisted textuality framework that integrates neural topic modeling with multimodal visualization to reveal latent thematic structures, intertextual linkages, and stylistic patterns, thereby enhancing analytical rigor for postgraduate literary research. Specific objectives are (1) to design a modular analytic pipeline that combines pre-trained contextualized embeddings (e.g., BERT-based representations) with dynamic topic models to generate interpretable thematic trajectories across serialized literary corpora; (2) to implement interactive visualizations that enable researchers to interrogate thematic clusters, co-occurrence networks, and stylistic features (e.g., lexical density, sentiment polarity) in relation to authorial voice and historical context; (3) to validate the framework through iterative case studies on canonical English-language fiction from the Victorian and Modernist periods, assessing coherence, interpretability, and scholarly utility; and (4) to evaluate the framework against established qualitative methods to determine its added value in generating novel research questions and supporting evidence-based argumentation. The methodology adopts a mixed-methods design grounded in computational linguistics and literary theory. The population comprises digitized English-language fiction texts from 1830–1930, totaling approximately 1.2 million tokens across 150 novels sourced from publicly accessible literary corpora and library-licensed datasets. A stratified sampling strategy selects 60 novels representing a spectrum of genres, authors, and publication years for in-depth analysis. Data collection employs (a) text preprocessing pipelines including tokenization, lemmatization, and sentence segmentation; (b) neural topic modeling using a neural variational document model augmented with contextualized word embeddings to produce dynamic topic distributions; and (c) extraction of stylistic metrics (lexical richness, syntax complexity, sentiment trajectories) and intertextual references. Instruments include a reproducible computational pipeline implemented in Python (PyTorch, Gensim, and Plotly) and a bespoke visualization dashboard facilitating interactive exploration. Validity and reliability are addressed through cross-method triangulation, parameter sensitivity analyses, and out-of-sample validation on two additional corpora (20 novels each). Data analysis proceeds in three stages. First, unsupervised modeling yields latent topic trajectories and topic-topic and text-topic associations, with coherence scores and perplexity tracked to ensure model stability. Second, a confirmatory phase tests whether computationally derived themes align with established literary categories via thematic mapping and expert evaluation by three literary scholars using a structured scoring rubric. Third, inferential analyses examine relationships between thematic emergence and textual features via regression analyses and time-series comparisons, supplemented by narrative case studies. The analytical framework is guided by theoretical perspectives from narratology and reception theory, with two explicit theoretical lenses Pratt’s concept of textual metaphorical networks and Ellison’s notion of dialogic intertextuality, adapted to computational discovery. A conceptual model illustrates how neural topic signals interact with visualization affordances to shape interpretive outcomes. Expected findings include (i) identification of coherent, shiftable thematic clusters that track genre- and period-specific concerns (e.g., urban modernity, social reform, psychological introspection); (ii) demonstration that visualization-enabled exploration reduces interpretation bias by making latent structures transparent and reproducible; (iii) evidence that the AI-assisted workflow can surface previously overlooked connections between authors and texts, contributing to new scholarly questions and debates; and (iv) validation that combining neural topic modeling with visualization yields richer, more defensible analytical narratives than traditional methods alone. Contributions to knowledge are twofold first, methodological, by offering a rigorously documented, reproducible AI-assisted toolkit for literary analysis that can be adapted to other languages and periods; second, theoretical, by providing empirical insights into how latent thematic structures relate to stylistic variation and intertextuality across canonical English fiction. The main conclusion anticipates that integrated neural topic modeling and visualization enhances both the efficiency and depth of literary analysis, enabling robust, scalable inquiry without sacrificing interpretive rigor. Recommendations include expanding the corpus to include non-canonical works and digital-born texts, refining interactive visualization to support collaborative authorship, and developing standardized evaluation protocols for AI-assisted literary analysis to promote broader scholarly adoption.