📄 Applying Lexical Constraints on Morpho-Syntactic Patterns for the Identification of Conceptual-Relational Content in Specialized Texts

Summary

The authors tackle a common weakness of surface-pattern extraction: frequent English prepositions such as for can signal several different conceptual relations. Their domain is the descriptions of wares and services in the Canadian Trademark Register, a corpus of about 12 million words that they treat as a sublanguage. The wider project supports classification under the Nice scheme, computer-assisted translation and trademark search.

They tag relations with triple labels and add Conceptual Constraint Variables (CCVs) to restrict when a pattern fires. CCVs are small named constraints that can be lexical, syntactic or morpho-syntactic. Each one either shapes the marker itself (endomorphic) or limits what surrounds it (contextual), and CCVs can be combined into larger patterns.

Applied to for, the constraints separate four readings: FUNCTION, DOMAIN, USER and a default SPEC. A manual evaluation shows very high precision for all four. Recall is also high, except for USER, which is limited by an incomplete lexicon.

The work shows that shallow, transparent mechanisms can produce reliable relational output without semantic dictionaries or unification grammars. The authors caveat that this works only within a narrow sublanguage.

Key points

• Corpus and task: the corpus is a roughly 12-million-word sample of the Canadian Trademark Register. Target relations are HYPONYMY, FUNCTION, COMPOSITION, DOMAIN, USER and NEGATION.

• Output format: relations are encoded as <HEAD_SUB_REL, FORM, RELATION> tags that wrap the marker, followed by the scope of the relation (the dependent text to its right).

• CCV types: constraints are lexical, syntactic (POS tags such as Penn Treebank NNS, or word order) or morpho-syntactic (word endings such as -ion, -ment, -ing). Each can be endomorphic or contextual, giving six variants.

• Readings of FOR in the data: FUNCTION 69.4%, SPEC 26.7%, DOMAIN 2.6%, USER 1.3%.

• Results by reading:

  Reading	Precision	Recall
  FUNCTION	99.4%	98.1%
  DOMAIN	100%	100%
  USER	100%	50%
  SPEC	90.3%	99.4%

  The weak USER recall comes from the difficulty of listing all user nouns.

Interesting ideas

• I1. Each pattern is split into a part describing optional and required words inside the marker (e.g. "especially designed for use in") and a separate condition on the material that follows it; this keeps many surface variants under one reading.
• I2. Word endings like -ion, -ment and -ing serve as a cheap signal that the complement names an action, which decides between a purpose reading and a beneficiary or target reading.
• I3. Constraints are small named units that refer to one another (a domain-word list reused inside "field of NP" and "NP industry" patterns), so every decision traces back to readable building blocks.
• I4. A vague default reading (SPEC, a plain restriction) is used when no specific pattern matches, instead of forcing a guess.
• I5. The evaluation reports how often each reading occurs alongside its precision and recall, which exposes that a rare class with perfect precision still misses half its cases because of lexicon gaps.

For Mova

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

• I1 (frames): prag could map oblique "for"-phrases to roles by keying on the head word's form and class, while treating "designed/intended for" variants as one marker, with the triggering condition logged as the explanation.
• I2 (world): when a "for"-object is a nominalization or gerund, the world builder could record a purpose or intended-event state rather than a recipient entity.
• I3 (learn): self-induced repair rules could be stored as composable named predicates, making them easy to audit, reuse and switch off.
• I4 (answer): readers could fall back to an underspecified relation and decline to answer purpose or user questions, rather than inventing a role.
• I5 (learn): per-role scores weighted by frequency would reveal which lexical lists (e.g. person nouns) need LLM-teacher expansion.