📄 ChiSense-12: An English Sense-Annotated Child-Directed Speech Corpus

Summary

Studies of how young children resolve word ambiguity in everyday speech have been limited because no sense-annotated corpus of child-directed speech existed with the needed detail. The authors built ChiSense-12 from CHILDES (version 2020.1), using 53 American and British English corpora of speech addressed to 958 children aged up to 59 months. They selected 12 ambiguous words, each with a frequent "dominant" sense and a rarer "subordinate" sense, and chose senses that earlier experiments showed preschoolers already understand. Every occurrence was annotated, with no downsampling. Each token received a sense label, and when the word served as a verb's object, the verb stem was also recorded. The final resource has 15,581 utterances. Because it captures which verbs combine with which senses, it supports separating two explanations for how children disambiguate: surface co-occurrence statistics and the semantic constraints verbs place on their objects (verb-event structure). The authors report preliminary experiments built on corpus statistics and plan to train Transformer-based models on the corpus. The corpus is freely available on GitLab.

Key points

• Source data: 72 British and American English corpora were downloaded from CHILDES through the R package childesr. The 53 corpora with target children up to 4 years (59 months) were kept, covering 958 children.
• Of 21,342 raw utterances containing the target words, 15,581 were retained (115,272 word tokens, 4,805 word types). On average the dominant sense accounts for 73% of uses (SD 13%).
• Eleven of the 12 words come from Rabagliati et al. (2013). The homophone pair flower/flour was added, and it is heavily skewed (3521 vs 350 uses, 91% dominant). Glasses is nearly balanced (683/620, 52%).
• The first author annotated each utterance with its sense and, for object uses, the preceding verb stem. Unclear cases were checked against neighbouring utterances and discarded if still unresolved. A second annotator validated 45 test conversations after training on 5.
• The authors contrast their work with Meylan et al. (2021), who used WordNet senses for CDI words and tagged only 50 tokens per 3-month interval for frequent words. That sampling thins out verb statistics for rare senses.

Interesting ideas

• I1. Annotating every token instead of a capped sample per frequent word keeps the co-occurrence profiles of rare senses intact. Capping quietly distorts exactly the minority cases you most want to study.
• I2. Each noun sense is stored together with the verb that takes it as an object. This turns a sense corpus into a resource of selectional preferences ("meet" favours a person reading of "star").
• I3. Test items are built from corpus statistics so that raw word association and verb plausibility point in opposite directions. The night/knight story is an example where the topic words favour the wrong reading.
• I4. The sense inventory is limited to readings the target audience is known to understand, using a coarse dominant/subordinate split instead of a fine-grained adult dictionary.
• I5. Annotators escalate to the surrounding discourse when a sentence alone is ambiguous, and drop the item if it stays unclear. Validators also get a third "something else" option, so they are never forced into a binary choice.

For Mova

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

• I1 (learn): When building silver data for sense or role statistics, avoid uniform per-lemma caps so that low-frequency readings keep reliable verb and argument distributions.
• I2 (frames): Use the UD obj relation plus VerbNet selectional restrictions to choose a noun's sense from its governing verb. The trace can show which verb constraint forced the choice.
• I3 (learn): Add negative-control items where bag-of-words association suggests the wrong answer, so evaluation confirms the structural path is actually doing the work.
• I4 (answer): For fables, tales and school tests, restrict ambiguous nouns to a small, audience-appropriate sense set. This reduces spurious WordNet distinctions that confuse the reader.
• I5 (world): Resolve entity senses by widening the window to earlier sentences, and record an explicit "unresolved/other" state instead of committing to a guess.