📄 Similarity between Words Computed by Spreading Activation on an English Dictionary

Summary

Kozima and Furugori address a gap in text analysis: computers can spot repeated words, but they struggle to detect semantic relatedness between different words, which is a key signal of lexical cohesion. The authors build a closed dictionary subsystem, Glossème, by keeping only the LDOCE entries whose headwords belong to the Longman Defining Vocabulary (LDV). Because LDV words are defined only with LDV words, the subsystem is self-contained. They convert it into a weighted semantic network, Paradigme, with forward links (referent) and backward links (référé). Similarity is computed by activating one word's node for a few steps, letting activity spread, and reading the activity that arrives at the other word's node. The result is a graded score in [0, 1] that separates related pairs such as cat/pet from unrelated ones such as cat/mat. The measure covers any LDV pair directly and other LDOCE words indirectly. The authors also use it for text similarity and coherence, and they are starting to apply it to text segmentation.

Key points

• Resource sizes: LDOCE has 56,000 headwords and LDV has 2,851 words. Glossème has 2,851 entries totalling 101,861 words, about 35.7 per entry. In Paradigme, the 2,851 nodes are joined by 295,914 links that carry no labels, roughly 104 for each node.
• Each node stores a headword, a word class, an activity value, a referent and a référé:
  • The referent is split into subreferents, one per numbered sense, and is described as intension.
  • The référé holds the nodes that point back to this node, described as extension.
• Link weights follow a fixed recipe:
  • The base weight is the inverse frequency of the word's root in Glossème.
  • The weight is doubled if the word sits in the head-part of a definition.
  • It is then shared among homograph nodes in proportion to their frequency.
  • Finally it is normalized to sum to 1. Earlier senses receive higher subreferent weight.
• Similarity is computed in three steps. First, all nodes are reset and w is activated with strength s(w) for 10 steps. Then the activity of w′ is read. That value is multiplied by s(w′). Example scores are σ(cat, pet) = 0.133722 and σ(cat, mat) = 0.002692.
• Known weaknesses come from the dictionary's style:
  • Definitions give sufficient conditions, not stereotypes, so tree/leaf scores only 0.008693.
  • Definitions are treated as word lists, which loses syntax. In the "lift" example, the sense of upward motion cannot be captured.

Interesting ideas

• I1. Using a dictionary's restricted defining vocabulary as a closed, self-defining core turns a sprawling lexicon into a compact graph, where every symbol is explained by other symbols in the same set.
• I2. Storing both outgoing definition links and incoming "who mentions me" links lets activation flow from what a word means (intension) and from what it is used to describe (extension).
• I3. The weighting scheme is simple and explainable: rare words count more, head-part words count double, ambiguous words are split across senses by frequency, and earlier senses are favoured.
• I4. Relatedness is graded by a bounded number of spreading steps instead of a binary same-category test, which also gives strength of relation, something thesaurus lookups lack.
• I5. The authors measure coherence and topic shifts from word-to-word cohesion, and they openly separate this from syntax-aware meaning. That gives a clear baseline to compare structured methods against.

For Mova

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

• I1 (world): The ontology layer could anchor WordNet/FrameNet concepts to a small closed defining core, so every concept bottoms out in traceable definitions rather than opaque IDs.
• I2 (world): The knowledge graph could keep explicit reverse edges with their own weights, so relatedness queries can explain whether a link came from definition or from usage.
• I3 (world): This recipe could weight edges built from WordNet glosses, so every relatedness score can be decomposed into human-readable factors.
• I4 (frames): Coreference and bridging (e.g., linking "pet" back to "cat") could use a short, step-limited activation score as an explainable cue for candidate antecedents.
• I5 (learn): A cohesion-only activation baseline could serve as a negative control, showing where frame-based understanding of fables beats association without structure.