📄 Learning the Countability of English Nouns from Corpus Data

Summary

Baldwin and Bond tackle automatic acquisition of noun countability: deciding, for each English noun type, whether it behaves as countable, uncountable, bipartite (scissors-like) or plural only. This lexical property affects parse interpretation, determiner and number choice in generation, and machine translation from languages such as Japanese that do not mark countability. The authors run a POS tagger, a chunker and a full parser over the written portion of the BNC. For every noun, they count how often it appears in countability-relevant constructions and turn those counts into a fixed-length vector organised into feature clusters. Gold labels come from merging COMLEX with the ALT-J/E transfer dictionary. A separate memory-based classifier for each class then labels unseen nouns, so a noun can receive more than one class. Measured against the merged lexicons, the predictions were correct 94.6% of the time. That is slightly above the agreement between the two dictionaries themselves and well above an all-countable baseline. The work shows that a deep lexical property can be learned from raw text cheaply and kept aligned with real corpus usage.

Key points

• The four categories being predicted are illustrated by dog for countable, furniture for uncountable, scissors (as in "a pair of scissors") for bipartite, and clothes or goods for plural only. Classes are predicted independently, so multi-membership is possible.
• Resources:
  • COMLEX 3.0 has about 22,000 noun entries, with 12,922 marked countable and 4,976 uncountable.
  • ALT-J/E gives 56,245 unique English entries.
  • The plural-only class was hand-checked down to 104 entries.
  • Corpus: BNC written component, about 90m w-units, with only its sentence tokenisation reused.
• Difficulty ceiling: COMLEX and ALT-J/E agree on 93.8% of class decisions. Nearly half of their disagreements involve nouns that ALT-J/E lists with two countabilities and COMLEX with one.
• Main result:
  • 4,982 nouns could be checked against the merged 63,581-noun lexicon.
  • On these, accuracy was 94.6%, against an all-countable baseline of 89.7% (a 47.6% error reduction).
  • Classifiers used the best-500 features, with no significant loss compared with the full feature set.
• Corpus-grounded check:
  • On 100 blind-annotated test nouns, the system matched human judgements 92.4% of the time, versus a baseline of 87.8%.
  • On a comparable sample of training nouns, the dictionaries themselves matched the human judgements only 86.8% of the time.
  • Errors stemmed mostly from the system outputting just one class for nouns that have two.

Interesting ideas

• I1. Estimate how hard a labelling task really is by measuring how often two independent gold lexicons agree, averaging per-class yes/no decisions (including shared negatives); this gives a realistic ceiling instead of pretending 100% is reachable.
• I2. Treat a lexical property as a type-level summary of token-level construction counts gathered with several pre-processing depths (tags, chunks, parses), rather than as a label on individual tokens.
• I3. Use one independent binary classifier per class instead of a single multi-way choice, so a noun like "beer" or "cake" can legitimately belong to both countable and uncountable.
• I4. Split the test items into frequency bands and report precision and recall per band; here this showed rare nouns get over-labelled countable while frequent ones get under-labelled.
• I5. Check a hand-built dictionary against blind, usage-based human annotation, which here revealed that the static lexicon matched real usage worse than the learned system did.

For Mova

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

• I1 (learn): Before training any lexical classifier, compare two sources such as WordNet-derived and Wikidata-derived labels to set an honest target and to flag the entries where disagreement is inherent.
• I2 (world): Mova can compile a countability table for its ontology by aggregating, over its own UD parses, how each noun combines with numerals, plural morphology and quantifiers like "much", with each entry traceable to the counted contexts.
• I3 (world): Entity typing in the text world (substance versus individual objects, needed for holdings and measures) could allow dual membership and let context decide which reading applies.
• I4 (learn): Mova's evaluation reports could bucket results by lemma frequency to catch biases that a single overall score hides, especially for silver-data-trained components.
• I5 (back): When generating text, determiner and plural choices could rely on corpus-attested countability statistics rather than only on dictionary flags, avoiding outputs like "two equipments".