📄 The Relationship between Semantic Similarity and Subcategorization Frames in English: A Stochastic Test Using ICE-GB and WordNet

Summary

The paper asks how strongly a verb's meaning predicts its syntactic argument patterns (subcategorization frames) in English. The authors turn this Levin-style intuition into a corpus-based statistical test. They start from frequent verbs in the Collins Cobuild English Dictionary that also appear in WordNet 2.1, giving 799 verbs. Pairwise similarity is computed with the Lesk measure from WordNet::Similarity. Bottom-up hierarchical clustering then builds a dendrogram, and a Z-score cutoff keeps only the tight clusters. For each clustered verb, frames are extracted from the ICE-GB treebank via ICECUP III, encoded as function+category labels. Each frame is weighted with a Resnik-style Kullback-Leibler selectional preference strength. Verbs are then graded by how well their dominant frames match those of their cluster-mates. Depending on how strictly agreement is defined, the hypothesis holds for about 42% to 73% of verbs. Agreement is markedly higher in the most tightly bound clusters. The work matters because it gives a reproducible, resource-driven way to quantify the link between verb semantics and syntax, rather than relying on intuition alone.

Key points

• Data pipeline:
  • About 1,600 top-frequency CCED words were intersected with WordNet verbs, leaving 799 verbs.
  • Clustering produced 61 significant clusters covering 157 verbs.
  • A cluster node counted as significant if its Z-score exceeded 1.
• Similarity and clustering:
  • Lesk (gloss overlap, including hypernym and hyponym glosses) was chosen because it yields larger values, which suits clustering.
  • At each step the most similar pair is merged, and the merged node's similarity to the other nodes is set to the minimum of its members' similarities.
• Frame extraction from ICE-GB:
  • Subjects and passive clauses were excluded.
  • Functional tags on relative pronouns (e.g. OD,NP) recover arguments that sit outside the relative clause.
  • Oblique complements are marked CO,PP.
  • Intransitive uses are encoded as "−".
• Results across 157 verbs:
  • Y (agreement >0.8): 66 verbs, 42.04%.
  • M (top two frames sum >0.5): 49 verbs, 31.21%.
  • N: 39 verbs, 24.84%.
  • ? (not found as a verb in ICE-GB): 3 verbs, 1.91%.
• Cluster tightness matters: the share of Y is about 63% in the 10 highest-Z clusters versus about 22% in the 10 lowest. The share of N is 11% versus 48%.

Interesting ideas

• I1. Representing a frame as a set of function+category pairs (e.g. "CO,PP / OD,NP") keeps grammatical role and phrase type together, so frames from different verbs become directly comparable signatures.
• I2. Ranking a verb's frames by KL-based selectional preference strength turns "which frames are licensed" into "how characteristic each frame is," which is more robust than raw counts or a yes/no lexicon.
• I3. Using cluster tightness (Z-score of the merge node) as a confidence signal works because tighter semantic clusters agree far more on syntax.
• I4. Dropping passives and subjects avoids biasing a verb's frame profile toward surface forms such as "be regarded."
• I5. Grading agreement in tiers (strong, partial, none, unseen) instead of a single accuracy number gives a more honest picture of how far a hypothesis holds.

For Mova

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

• I1 (frames): Frame signatures harvested from UD trees could be keyed as relation+category pairs and compared against VerbNet class expectations to flag unusual role assignments.
• I2 (world): The compiled verb-class knowledge could store a strength score per verb–frame pair, letting the frame builder prefer characteristic readings when parses are ambiguous.
• I3 (learn): Silver frame data could be trusted more for verbs in tight WordNet clusters and checked more for verbs in loose ones, serving as a cheap prior when selecting training examples.
• I4 (learn): When inducing frame statistics from treebanks, voice should be normalized or passives filtered so that alternations do not distort a verb's canonical frame distribution.
• I5 (answer): Evaluations of verb-class predictions could report the same tiers, separating "unseen" cases from genuine failures, which acts as a simple negative control.