📄 The Pilot Corpus of the English Semantic Sketches

Summary

The paper tackles a weakness of classic word sketches, such as those in Sketch Engine. These group a verb's dependents by syntactic relation and blend all of the verb's senses into one picture. The authors build semantic sketches instead. For each separate verb sense, a sketch lists the semantic roles the verb takes and ranks the most typical words filling each role. The roles and senses come from the Compreno parser's full semantic analysis, so a sketch reflects neighbours in the parse graph rather than in a linear word window. The new English pilot corpus contains 100 hand-selected sketches built from a mixed-genre corpus, and each is paired with a Russian sketch for the same semantic class. The authors use these pairs for contrastive analysis, for example how English pour divides its objects compared with Russian лить/сыпать. They also classify the typical defects of automatically built sketches. The resource gives an interpretable, sense-aware view of verb meaning that can support word sense disambiguation (WSD) and cross-language lexicography.

Key points

• English corpus: 100 manually vetted sketches, built from technical, news and fiction texts. These texts contain about 14 million verb-headed syntactic links.
• Pairing and verb selection: The English sketches are matched with 84 Russian sketches, giving 113 English–Russian pairs. Verbs were drawn from the same semantic classes as the earlier polysemous Russian corpus.
• Frequency thresholds: An English verb needed at least 200 links. The Russian threshold was 2000, on a dataset of more than 36 million links.
• Display and data source: Each role normally shows its eight most frequent fillers. Senses and roles come from Compreno's thesaurus-like semantic hierarchy.
• Error types identified:
  • sparse ("empty") role columns;
  • slots belonging to a different sense;
  • wrong fillers caused by frequent homonyms outscoring correct analyses;
  • preposition ambiguity, e.g. a for-phrase expressing distance misfiled as Purpose;
  • statistical fluctuations between languages.
• Availability: The corpus is on GitHub, with plans to grow it and add more languages.

Interesting ideas

• I1. Build a separate role-by-filler profile for each verb sense rather than for each lemma, so that typical arguments describe a meaning instead of a word form.
• I2. Take a verb's context from its typed dependency neighbours in the parse graph instead of a fixed linear window. This keeps the representation interpretable and filters out irrelevant nearby tokens.
• I3. Read aggregated sketches as a diagnostic of parser behaviour, since misplaced fillers expose systematic role-labelling mistakes. An example is a distance phrase landing under Purpose.
• I4. Separate truly narrow slots from sparse data: verbs acting like lexical functions legitimately have very few fillers, while other gaps simply reflect too little data.
• I5. Compare same-sense verbs across languages by the semantic features of their object fillers, as with liquid versus granular objects for pour compared with лить/сыпать.

For Mova

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

• I1 (world): Compiling such per-sense filler tables from parsed text could give verb classes explicit selectional preferences that the world layer can inspect and cite.
• I2 (frames): Role assignment could check candidate arguments against filler statistics gathered over UD/frame edges, staying fully explainable with no bag of words.
• I3 (learn): Scanning compiled role/filler tables for outliers could surface candidate self-induced repair rules for frequent misclassifications.
• I4 (world): Marking each slot as "closed by nature" or "under-observed" would prevent the knowledge base from drawing strong conclusions from thin evidence.
• I5 (back): Translation through the meta-world could choose the target verb from the features of the object entity rather than from a one-to-one lexical mapping.