📄 Reinforcing English Countability Prediction with One Countability per Discourse Property

Summary

The paper tackles predicting whether an English noun occurrence is countable or uncountable without using articles or number marking. That constraint matters for spotting determiner and plural errors in learner writing. The authors propose "one countability per discourse", modelled on Gale et al.'s one-sense-per-discourse idea: repeated mentions of a noun within one text tend to share countability.

A pilot test on the British National Corpus showed the per-text majority label beats a corpus-wide majority baseline by about 10 points. The property is still clearly weaker than its word-sense counterpart.

The method works in two passes:

• A base maximum entropy classifier first labels every occurrence in a text, keeping only confident predictions to estimate the text's majority countability.
• That estimate is then added as one more input feature, so the learner decides how much to trust it.

Training labels come automatically from heuristic tagging rules. Across 23 dual-countability nouns, this reinforcement gives a small but statistically significant gain over the plain classifier. Bluntly overwriting predictions with the majority label actually hurts.

Key points

• The data comes from the British National Corpus's written-text portion, with each text treated as one discourse. 314 texts (about 10%) were held out for testing, and the remainder was auto-labelled for training with the tagging rules of Nagata et al. (2005b). The OAK system was used to extract noun phrases and heads.
• There are 23 target nouns, taken from Huddleston and Pullum (2002), that can be both countable and uncountable (e.g. paper, advantage, hair, work).
• In the pilot test, the discourse-majority label (MCD) averaged 0.851 accuracy against a 0.754 corpus-majority baseline. The size of this effect differed a lot between nouns: it was strong for aid and cover, and weak for advantage and improvement.
• Average accuracy across the methods:
  • Proposed: 0.857
  • ME alone (opennlp.maxent): 0.842
  • ME+MCD (overwrite with majority): 0.828
  • Decision list (DL): 0.812
• The gain over ME is significant at 99% (paired t-test). The authors regard the 1.3–1.7% improvements reported for one sense per discourse as a rough upper bound for this kind of gain.

Interesting ideas

• I1. Instead of forcing a discourse-level consistency constraint, the paper encodes it as a soft feature and lets the learner weigh it against local context, which protects exceptions like the idiom "take advantage of".
• I2. When the majority label is estimated from the model's own predictions, low-confidence outputs are dropped first, so spurious guesses do not pollute the aggregate.
• I3. Texts where the noun appears only once trivially satisfy the consistency property, so the authors mark their majority value as "unknown" to keep the model from over-trusting the feature.
• I4. Training labels are generated by a small decision tree of surface rules (plural form, indefinite article, words like much or another), and instances the rules cannot classify are discarded.
• I5. The evaluation includes a naive "overwrite everything with the majority" variant as a negative control, and plots per-item accuracy against the diagonal; this shows that the raw signal helps while the naive use of it hurts.

For Mova

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

• I1 (frames): Coreference chains could add a "majority reading of this entity across the text" feature for sense or role decisions, rather than a hard rule that might break on idioms.
• I2 (learn): When pooling silver labels from witness parsers per document, only high-agreement or high-confidence votes should count toward document-level priors.
• I3 (learn): Any self-derived aggregate feature should get an explicit "no evidence" value for singleton cases, so training does not learn a degenerate shortcut.
• I4 (world): Mass/count status of entities, which feeds quantities and holdings, could be bootstrapped from such cue rules over large corpora, with an explicit "?" class that is dropped from training.
• I5 (learn): Each new consistency heuristic should be tested against its naive hard-override version, with per-lemma scatter plots, to catch side effects that averages hide.