📄 A database of semantic clusters of verb usages

Summary

The paper presents VPS-30-En, a small, freely available lexical resource covering 30 English verbs, including access, claim, say, tell, throw and yield. It is a revised snapshot of Patrick Hanks' Pattern Dictionary of English Verbs (PDEV), which describes how verbs are used in the BNC through Corpus Pattern Analysis (CPA). In CPA, each verb sense is a pattern whose argument slots are filled by semantic types or lexical sets, paired with an "implicature" that paraphrases its meaning.

The authors wanted gold-standard data for automatic clustering of verb usages. They wrote explicit annotation guidelines, cleaned the entries and their annotated concordances, extended the slot scheme with more syntactic detail, and ran several multi-annotator rounds. They then analysed where annotators disagreed. Two major sources of disagreement are rooted in the language itself rather than in bad pattern design:

• participial forms that blur passive and adjectival readings;
• semantic modulation of nouns.

When these are discounted, agreement rises substantially. The work matters because it tests whether pattern-based verb sense distinctions are reliable enough to learn automatically.

Key points

• PDEV is built on BNC50, a written subset of the BNC with roughly 50 million tokens. It has 694 completed verbs with 2,663 patterns, uses about 200 semantic types, and covers only about 10 % of verb occurrences.
• Concordance labels in the samples:
  • a pattern number marks a prototypical use;
  • "number.e" marks an exploitation (figurative or syntactically deviant use);
  • "u" marks a sensible use that fits no pattern;
  • "x" marks noise such as meta uses, named entities or mistagged words.
• Reference samples usually contain 250 lexicographer-annotated lines per verb, or all occurrences for rarer verbs.
• Pairwise Cohen's kappa gives nearly the same baseline as Fleiss' kappa. When exploitation marks and participle-pattern disagreements are ignored, 20 verbs exceed 0.8, 6 exceed 0.7, and the rest stay above 0.6.
• Participle-driven disagreement affects 14 verbs, including enlarge, cool, crush, swell and wake.

Interesting ideas

• I1. Agreement can be recomputed with chosen disagreement classes neutralised (exploitation marks within one pattern, participle versus transitive), which separates inherent textual vagueness from flawed category design.
• I2. Every usage gets a graded fit label: prototypical, exploitation of a known pattern, unmatched but valid, or not a real verb use. This is more informative than a forced single sense.
• I3. Patterns pair typed argument slots (semantic types or lexical sets of collocates) with an implicature that mirrors those slots. Each sense therefore carries its own meaning paraphrase.
• I4. Arguments are classified by their function, whatever form they take (noun or clause), and slot categories are deliberately aligned with what parsers actually output. For example, prepositional objects are treated as adverbials.
• I5. Participles lie on a continuum from passive to adjective, where the agent may be underspecified, irrelevant or absent. The authors therefore treat the participial and the causative categories as two non-exclusive perspectives on one event.

For Mova

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

• I1 (learn): When witness parsers or teacher tiers disagree on silver data, we could tag the disagreement type and score agreement with and without each type, so harmless vagueness is not treated as a parser error.
• I2 (frames): The frame matcher could output the same four-way verdict, so figurative or deviant uses stay explainable instead of being silently forced into the closest VerbNet frame.
• I3 (world): Verb-class knowledge could store slot-mirroring paraphrases per pattern, and these could later drive explanations in answers and regeneration in the back stage.
• I4 (tree): The UD-to-frame mapping could adopt this rule, defining frame slots over dependency functions that our parsers produce reliably rather than over hard-to-parse distinctions.
• I5 (world): Forms like "was enlarged" could be recorded as a resulting state with an optional, possibly unknown causer, instead of committing to a guessed agent.