Text Mining Freud’s Complete Works: A Lexical-Affective Cartography across Volumes

Technical Summary: Text Mining Freud's Complete Works

Problem Statement
Sigmund Freud's theoretical corpus is a foundational yet complex body of work in the humanities and health sciences. Its conceptual density and vast scope (spanning over four decades) make it difficult to identify large-scale patterns, structural regularities, and longitudinal shifts in conceptual emphasis through traditional close reading alone. While hermeneutic and clinical approaches are indispensable, they face limitations when analyzing global patterns across thousands of pages. This study addresses the need for a macrotextual perspective to investigate the lexical, conceptual, and affective organization of Freud's Complete Works without replacing traditional interpretation, but rather expanding the scale of observation.

Methodology
The study employed an exploratory text-mining approach on the Portuguese edition of Freud's Complete Works (Editora Imago), comprising 23 editorial volumes published between 1886 and 1939.

• Corpus Construction & Preprocessing: The raw PDF files were converted to plain text using the pdftools package in R. A rigorous curation process removed editorial paratexts, headers, footers, metadata, and conversion noise, retaining 12,524,088 characters across 4,145 pages. Texts were processed according to "tidy text" principles, including lowercasing, punctuation removal, and harmonization of lexical variants (e.g., singular/plural forms of key concepts) to support concept-level analysis.
• Lexical Analysis: The study calculated absolute and relative word frequencies, extracted bigrams and trigrams, and constructed directed lexical networks to visualize recurrent associations. A manually curated matrix of psychoanalytic keywords (e.g., dream, sexuality, symptom, unconscious, repression) was tracked longitudinally across the 23 volumes.
• Affective Analysis:
  • Polarity: Lexical polarity was assessed using the SentiLex sentiment lexicon for Portuguese.
  • Discrete Emotions: The NRC Emotion Lexicon (EmoLex) was used to map words to eight basic emotions (anger, anticipation, disgust, fear, joy, sadness, surprise, trust) and two polarity categories.
  • Normalization: Emotional proportions were normalized using z-scores within each category to highlight volume-specific variations relative to the corpus mean.
• Statistical & Multivariate Analysis:
  • Correlations: Spearman's rho correlations were computed between the relative frequencies of psychoanalytic keywords and NRC emotion categories across volumes.
  • Contextual Enrichment: An over-representation analysis using the hypergeometric test (with Benjamini-Hochberg correction) determined if specific emotion categories were statistically enriched in the lexical context surrounding key psychoanalytic terms.
  • Principal Component Analysis (PCA): Two PCA models were applied: one to cluster conceptual keywords based on their associated affective profiles, and another to map the 23 editorial volumes in an affective-lexical space based on normalized emotion proportions.

Key Results

1. Lexical Architecture: Computational analysis successfully recovered recognizable conceptual nuclei of Freudian theory, including dream, sexuality, symptom, neurosis, hysteria, unconscious, desire, pleasure, reality, and childhood. Directed bigram networks revealed organized chains of association (e.g., "dream interpretation," "sexual life").
2. Longitudinal Variability: Key concepts did not appear homogeneously across the collection. Instead, their relative intensities oscillated across editorial volumes, indicating shifts in thematic focus consistent with the historical development of psychoanalysis.
3. Affective Profiles: The corpus exhibited a dynamic affective landscape rather than a uniform tone. Longitudinal analysis revealed oscillations in polarity and multidimensional emotional profiles involving trust, anticipation, fear, sadness, anger, disgust, joy, and surprise.
4. Affective-Semantic Clustering:
   • Keyword Correlation: Clinical-symptomatic terms (symptom, neurosis, hysteria, anxiety) showed strong associations with negative affective categories (fear, sadness, negativity). In contrast, interpretive and metapsychological terms (dream, desire, pleasure, unconscious, censorship) clustered near dimensions of trust, positivity, surprise, and joy.
   • Enrichment: Contextual enrichment analysis confirmed that clinical terms were surrounded by words enriched for negative emotions, while dream-related terms were associated with elaborative and interpretive affective fields.
   • PCA of Concepts: The first principal component (PC1) clearly separated clinical-symptomatic terms from dream-interpretive and metapsychological terms, explaining 64.9% of the variance.
5. Volume Cartography: PCA applied to the editorial volumes revealed that the collection is not affectively homogeneous. Volumes were ordered along gradients distinguishing texts with higher negative loads (fear, sadness) from those with higher positive or elaborative loads (joy, trust, anticipation), suggesting distinct "affective signatures" for different editorial moments.

Key Contributions

• Methodological Integration: The study demonstrates the viability of combining distant reading, text mining, and lexicon-based emotion analysis to investigate classical psychoanalytic corpora. It functions as a complementary analytical lens to traditional hermeneutics.
• Empirical Validation: The computational approach successfully recovered the expected theoretical architecture of Freud's work, serving as an internal validation of the methodology.
• Affective-Conceptual Mapping: The research provides quantitative evidence that Freudian concepts are not randomly distributed but are organized into distinct affective-semantic fields (clinical-symptomatic vs. dream-interpretive/metapsychological).
• Longitudinal Cartography: The study establishes that Freud's work exhibits a measurable, longitudinal lexical-affective organization, where conceptual shifts are accompanied by changes in the affective texture of the vocabulary.

Significance and Claims
The paper claims that Freud's work, as represented in the analyzed Portuguese edition, possesses a "measurable longitudinal lexical-affective organization." The study does not propose that these findings replace clinical or theoretical interpretation; rather, it argues that computational analysis offers an "additional layer of quantitative evidence" that helps map how concepts and affective vocabularies are articulated throughout Freud's production.

The authors position this work as an "exploratory computational cartography" that complements traditional interpretation. By revealing structural patterns and affective signatures that are difficult to detect through close reading alone, the study opens new possibilities for digital humanities approaches to psychoanalysis. It suggests that psychoanalytic corpora can be investigated in a reproducible, comparative, and theoretically oriented manner, allowing for the formulation of refined hypotheses regarding the relationship between language, affect, and theoretical construction.

Limitations Acknowledged
The authors explicitly note that lexicon-based analyses cannot capture irony, ambivalence, or the specific conceptual density of terms within a theoretical tradition. Furthermore, the analysis is based on the Portuguese translation, meaning the patterns identified are mediated by translation and editorial decisions, not a direct reconstruction of Freud's original German. The use of editorial volumes as units may also smooth over internal variations within specific texts. Consequently, the findings are presented as a starting point for "augmented reading" and future granular analyses rather than a definitive interpretation of Freud's theoretical evolution.