📄 The Clustering of Author's Texts of English Fiction in the Vector Space of Semantic Fields

Summary

Clustering documents by raw word frequencies produces vectors as large as the vocabulary, which makes the computation heavy. Pavlyshenko instead describes each document by how often its words fall into broad semantic fields, which are word groups that share a general concept such as motion or perception. The fields come from WordNet's lexicographer files for nouns and verbs. Each field's weight is the summed frequency of its member words, normalized so that every document's profile adds up to one. The study clusters 503 English fiction works by 17 authors hierarchically, using Ward's method and Euclidean distance, and stops at 20 clusters. A second run first projects the field-by-document matrix onto its leading singular vectors (SVD, as in LSA) and then clusters again. Some clusters mix many writers, while others are dominated by one or a few. The author interprets the dominated clusters as regions of semantic space that reflect individual idiolects. This pattern persists at 10 dimensions but disappears at 3. The result suggests that a small, interpretable feature space can still carry authorial signal.

Key points

• The corpus contains 503 literary works by 17 authors of English fiction.
• The features come from WordNet lexicographer files:
  • 26 noun files (e.g. noun.feeling, noun.artifact), from which 54,464 words were selected;
  • 15 verb files (e.g. verb.motion, verb.perception), from which 9,097 words were selected;
  • derived word forms were also counted toward their fields.
• Clustering is agglomerative with Ward's criterion and Euclidean distance, halted when 20 clusters remain. Results are shown as a dendrogram and per-cluster histograms of authors.
• SVD of the field-frequency matrix shows singular values falling off sharply. The first 10 singular directions form an orthogonal space of "secondary" fields.
• Author-dominated clusters appear in both the full field space and the 10-dimensional SVD space, but not in a 3-dimensional one. Evaluation is visual only; no purity or accuracy figures are reported.

Interesting ideas

• I1. WordNet supersense (lexicographer-file) distributions give a document a compact profile in which every dimension has a human-readable name, unlike raw vocabulary vectors or opaque embeddings.
• I2. Keeping noun fields and verb fields as separate axes splits "what things a text talks about" from "what kinds of actions happen" while keeping both coarse.
• I3. Running SVD on an already small, meaningful matrix and sweeping the target rank (10 keeps the effect, 3 destroys it) locates where the signal actually lives.
• I4. Inspecting the label composition of each cluster separates regions of the space that discriminate between classes from regions that are semantically generic and uninformative.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (world): Compute a supersense profile over the entities and events already placed in a text's world, not over tokens, to give an explainable genre or topic signature that can steer which compiled knowledge to load.
• I2 (tokens): Attach the lexfile supersense to each lemma in crate en as a cheap semantic tag that later frame and role assignment can consult or check against.
• I3 (learn): Use rank or feature-ablation sweeps as an evaluation habit, treating the setting where a claimed effect vanishes as a built-in negative control.
• I4 (learn): Cluster sentences by supersense profile and overlay witness-parser disagreement rates, so repair-rule induction and LLM-teacher effort target the clusters where errors concentrate.