📄 A sense-based lexicon of count and mass expressions: The Bochum English Countability Lexicon

Summary

Work on the count/mass distinction usually labels whole lemmas, treats countability as a yes/no property, and relies on a handful of textbook nouns. This paper presents BECL 2.1, a public lexicon that instead assigns countability to individual noun senses. Lemmas were taken from Open ANC (nouns occurring more than 10 times) and paired with their WordNet senses, which gives about 11,800 annotated noun-sense pairs.

Annotators never judged countability directly. They placed each pair into a few syntactic frames (for example with "more", in the plural, or as a definition subject with or without an article) and answered follow-up semantic questions. The combinations of answers define 18 fine-grained classes, grouped into four major ones: count, mass, both, and neither.

The release also contains:

• WordNet similarity sets based on LCH,
• flags for proper names,
• flags for orthographic variants,
• a list of nouns whose senses fall into different classes.

A first corpus study of mass nouns sorts their count-like uses into recurring shift types. The resource makes it possible to separate true dual-life senses from plain polysemy.

Key points

• BECL 2.1 has 11,875 noun-sense pairs: 8,434 count, 2,427 mass, 699 both, and 315 neither, spread over 18 subclasses. It is released as CSV tables at count-and-mass.org.
• The four tests Syn1–Syn4 and the two semantic tests Sem1–Sem2 allow 80 logically possible classes. Only 18 of these are attested with unanimous judgements from at least two annotators.
• Similar senses were computed with the NLTK WordNet module using the Leacock–Chodorow (LCH) score, with a fairly strict cutoff of 2.5.
• Two annotators marked proper names, organisations and locations, reaching Krippendorff's α = 0.82. Most proper names are "neither", but some are count, mass or both (e.g. voodoo, coke, exec).
• The authors parsed corpus sentences for regular-mass (class 528) lemmas, using a parser identified by a footnote link to the Stanford nndep package. After filtering, 374 lemmas in 5,536 sentences showed mass-to-count shifts:
  • about 25% are kind-of readings,
  • about 15% are event or act instantiations,
  • a further type reads the noun as a unit or portion of a substance.

Interesting ideas

• I1. Annotating at the noun-sense level separates "dual life" senses (the same sense is both count and mass) from "multiples" (different senses of one form land in different classes), a split that form-level lexicons cannot make.
• I2. Rather than asking annotators an abstract question ("is it mass?"), the authors had them test the word in concrete diagnostic frames and then derived the class from the pattern of answers, which makes the labels reproducible and auditable.
• I3. Grouping near-synonymous senses with an LCH threshold surfaces synonym pairs whose countability differs (pay vs. salary, poison vs. toxicant), pointing to cases where meaning alone does not predict grammar.
• I4. Corpus uses of strictly mass nouns in count-like syntax were sorted into a small set of shift types: a portion or unit of a substance, a kind of something, and an event or act instance.
• I5. Proper names were flagged separately and turned out not to be uniformly "neither", so they need their own countability handling.

For Mova

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

• I1 (world): Attach countability to WordNet synsets in the ontology, so that resolving the sense of "matter" also tells the world model whether to create countable individuals or a quantity of stuff.
• I2 (learn): Build LLM-teacher or silver-data labelling as a battery of frame-insertion tests whose answer vector decides the label, so every assigned class comes with an explainable justification.
• I3 (back): During generation, choose determiners and number per lemma from the lexicon instead of per concept, so paraphrasing "salary" as "pay" does not keep an article or plural that no longer fits.
• I4 (frames): When a lexicon-mass noun appears pluralised or with "a", flag it as coercion and choose among the unit, kind and event readings, rather than treating it as a tagging error.
• I5 (tokens): Keep a proper-name homograph flag next to the lemma (e.g. "coke") so the tagger and the downstream mass/count logic can switch between the brand reading and the substance reading.