Joyce-aware, linguistics-led extraction and analysis of similes (and near-similes) across Dubliners—built for reproducibility, auditability, and close reading.
This repository detects and categorises similes in Dubliners using weak supervision and transparent models, aligned to Comparative Suspension Theory (CST): Standard, Quasi, Quasi-Fuzzy, Framed, and Silent.
Much NLP work collapses “style” into generic signals. This project starts from linguistics and literary scholarship: it encodes Joyce-specific comparators and constructions, generates silver labels with rule functions, trains simple models for legibility, and surfaces errors as reading prompts rather than hidden noise. The goal is a method you can rerun, inspect, and argue with.
- Notebooks / scripts for:
- text loading and sentence segmentation;
- Joyce-aware rule functions for comparators and constructions;
- weak labelling to create silver data from curated seeds;
- featureisation and transparent baselines (LinearSVC / LogisticRegression);
- reports: concordances, confusion sets, and CST distribution tables.
- Standard — Explicit, local comparator (
as,like,as if) with governor–target contained within a short span (typically a clause). - Quasi — Attenuated or approximate comparison (hedges like “sort of / kind of / almost,” or weakened comparator semantics).
- Quasi-Fuzzy — Deliberately vague or open frames in which the compared domains are only loosely specified.
- Framed — Extended-metaphor behaviour realised as simile: a comparison is introduced and then sustained across a longer span by framing devices (punctuation/phrasing), often with list-like amplifications, resumptive patterns, or re-entry of the comparator. Typical surface cues include:
- colon/semicolon/dash frames that open a descriptive field (e.g., “— like …, like …; like … —”),
- parentheticals or appositives that carry forward the comparison,
- anaphoric/resumptive comparators (“as if … as if …”) across multiple clauses. The key property is distributed realisation: the simile is not a single local unit but a span with one or more comparator anchors and a framing structure that keeps the comparison alive.
- Silent — Comparison is implied with no overt comparator; syntax/semantics make the comparative force recoverable.
In the weak-labelling step, Framed is detected when the comparison is introduced and then sustained across a longer span using framing punctuation or resumptive comparator patterns (e.g., repeated “like / as if”, parentheticals, colon/semicolon/dash frames). Conflicts with Standard are resolved in favour of Framed when re-entry or framing is present and span length/comma density indicate distributed elaboration.
Biehle, M. (2025). joyce-dubliners-similes-analysis (Version 2025.10.26). GitHub repository: https://github.com/mahb97/joyce-dubliners-similes-analysis
# Python 3.10+
pip install -U pandas numpy scikit-learn nltk spacy beautifulsoup4 tqdm jupyter
python -m spacy download en_core_web_sm