Eigenmood Space: Uncertainty-Aware Spectral Graph Analysis of Psychological Patterns in Classical Persian Poetry

Technical Deep Dive: Eigenmood Space

Overview

This work presents an uncertainty-aware computational framework for analyzing psychological patterns in classical Persian poetry at scale. The key contributions are:

• Developing a multi-label psychological ontology to map individual poetic verses to recurring affective configurations.
• Aggregating verse-level annotations into a poet-level "Eigenmood" representation that captures both distributional prevalence and relational structure.
• Quantifying "poetic individuality" as divergence from a global baseline, while preserving uncertainty through abstention and confidence weighting.
• Validating the framework against human judgments and demonstrating its ability to support both distant reading summaries and targeted close reading.

Problem & Context

Classical Persian poetry is a sustained archive of affective expression through metaphor, convention, and rhetorical indirection. This makes close reading indispensable but limits reproducible comparison at scale. The authors introduce an uncertainty-aware computational approach to model psychological patterns across tens of thousands of verses.

Methodology

• Verse-level annotations: Annotate each verse with a set of psychological concepts (e.g., melancholia, romantic obsession) along with confidence scores and abstention flags.
• Poet-level aggregation: Aggregate the verse-level annotations into a Poet × Concept matrix, representing each poet as a probability distribution over concepts.
• Individuality quantification: Measure the divergence of each poet's distribution from a global baseline using Jensen-Shannon divergence and Kullback-Leibler divergence.
• Relational structure modeling: Build a confidence-weighted concept co-occurrence graph and derive an "Eigenmood" embedding through Laplacian spectral decomposition.

Data & Experimental Setup

• Corpus: 61,573 verses across 10 poets, drawn from the Ganjoor digital corpus.
• Annotation ontology: 9 psychological constructs, including melancholia, romantic obsession, and spiritual narcissism.
• Validation: A stratified sample of 500 verses annotated by two independent human experts to assess label reliability, abstention appropriateness, and confidence calibration.

Results

• Substantial uncertainty: 22.2% of verses are abstained, indicating the analytical importance of uncertainty modeling.
• Heterogeneous individuality: Poets exhibit varying degrees of divergence from the global baseline, ranging from Khayyam and Parvin as high-distance outliers to Hafez and Shahriar as baseline-adjacent.
• Relational structure matters: The Eigenmood embedding captures directional contrasts in concept coupling that are not apparent from marginal prevalence alone.

Interpretation

• The profiles model recurring rhetorical-affective configurations in poetic discourse, not clinical diagnoses of historical individuals.
• Divergence and Eigenmood patterns support probabilistic cultural-psychological modeling, identifying differential tendencies across poets without collapsing the tradition into monocausal claims.
• The framework combines distant reading summaries with targeted close reading, providing both macro-scale comparisons and verse-level exemplars.

Limitations & Uncertainties

• The psychological ontology is an operational construct, not a latent clinical taxonomy. Its validity depends on prompt design, model training data, and cultural-semantic fit.
• Abstention is not a nuisance but a substantive property, introducing selection bias that must be diagnosed and reported.
• Small subcorpora (e.g., Khayyam) require careful interpretation of point estimates due to finite sample effects.

What Comes Next

• Exploring alternative graph constructions, richer metadata, and cross-corpus transfer to other literary traditions.
• Expanding the evidentiary horizon while preserving interpretive caution, enabling robust comparative claims and transparent pathways from macro structure to inspectable verse evidence.