Abstract

This thesis advances the computational understanding and manipulation of textstyles through three interconnected pillars: (1) Text Style Transfer (TST),which alters stylistic properties (e.g., sentiment, formality) while preservingcontent; (2)Authorship Attribution (AA), identifying the author of a text viastylistic fingerprints; and (3) Authorship Verification (AV), determiningwhether two texts share the same authorship. We address critical challenges inthese areas by leveraging parameter-efficient adaptation of large languagemodels (LLMs), contrastive disentanglement of stylistic features, andinstruction-based fine-tuning for explainable verification.

 

Quick Read (beta)