📄 An annotated English child language database

Summary

Child-language studies increasingly rely on large naturalistic corpora such as CHILDES, but its English material mostly carries only POS tags, lemmas and parses. This makes it hard to pose questions that cut across several kinds of information. The authors build the English CHILDES Verb Database (ECVD), which places the English CHILDES transcripts in a searchable, modular database. Each sentence carries analyses from three sources:

• the MEGRASP dependency output shipped with CHILDES,
• RASP trees and grammatical relations,
• the hand-curated CHILDES Treebank for the Adam corpus.

Verbs are tagged with Levin's semantic classes. Verbs and nouns also receive psycholinguistic attributes, such as age of acquisition, familiarity, concreteness, imageability, frequency and syllable count, drawn partly from the MRC Psycholinguistic Database. Sentences are indexed by the verbs they contain. Queries can therefore combine a lemma, a grammatical relation and a verb class. A comments field records failed or near-matching alignments between parsers.

The result is an integrated resource for studying how the language heard and produced by children grows in complexity with age, which bears on poverty-of-the-stimulus debates.

Key points

• The database covers roughly 4.84 million raw sentences, with differing portions parsed by MEGRASP, RASP and the CHILDES Treebank.
• In a gold-standard evaluation, MEGRASP found 96% of the correct dependencies and assigned the right GR label 94% of the time.
• The CHILDES Treebank part covers Brown's Adam corpus. It was parsed with the Charniak parser in Penn Treebank style and then hand-curated, following Pearl and Sprouse (2012).
• Verb annotation uses Levin's (1993) 190 fine-grained subclasses, which span about 3,100 verb types.
• The MRC database has 150,837 entries with 26 properties, but coverage is uneven: imageability, for example, exists for only 9,240 words.

Interesting ideas

• I1. Several parsers' outputs are stored side by side for the same sentence, so their differing precision and recall can be exploited much like annotator agreement. A query can also be routed to whichever source is most reliable for that pattern.
• I2. Sentences are indexed by the verb classes of their verbs, so one class query (e.g. verbs of running) also retrieves slide, roll and stroll without listing them.
• I3. An explicit comments field logs where sentence alignment between pipelines failed or only nearly matched, typically because of non-words, metacharacters and punctuation handled differently.
• I4. Lexical entries are enriched with graded psycholinguistic attributes such as age of acquisition, imageability and familiarity, alongside syntactic data.
• I5. The database is designed to be modular, so new annotation layers can be added without rebuilding existing ones.

For Mova

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

• I1 (learn): Keep each witness parser's tree per sentence and record which witness proves most accurate for each construction, then weight silver-data votes by construction rather than globally.
• I2 (world): Index compiled knowledge and corpus examples by VerbNet/Levin class so frame lookups and example retrieval generalise across class members instead of single lemmas.
• I3 (tree): When converting or aligning treebanks across UD dialects, store per-sentence mismatch notes on tokenisation differences so conversion errors stay visible and auditable.
• I4 (answer): Use age-of-acquisition and familiarity scores to estimate how hard a school-test question or fable's vocabulary is, and to pick simpler paraphrases when explaining answers.
• I5 (frames): Store pragmatic layers such as speech acts, coreference and implicature as separate attachable annotation tiers over the same sentence IDs, so each can be revised independently and remain explainable.