📄 Annotating progressive aspect constructions in the spoken section of the British National Corpus

Summary

Parsers trained on written English often perform poorly on speech, and spoken data with expert linguistic labels is scarce. Caines and Buttery release a set of manual annotations for every sentence in the spoken part of the British National Corpus (sBNC) that contains a progressive aspect verb group. This covers roughly ninety thousand of about one million sentences, or 93,253 verb groups. The work began as a study of "zero auxiliary" constructions, such as a question asked without the expected "are". Each verb group is labeled for seven form-based properties: register spontaneity, subject type, subject person-number, auxiliary form, clause type, tense and polarity. Because the BNC license forbids redistributing its text, the release is stand-off XML with a DTD. Users link each record to their own BNC copy through document and sentence identifiers. A single annotator did all the labeling, and a later re-check of a sample found over 99% agreement. The counts show that dropped auxiliaries cluster in subjectless progressives and in second-person questions. The resource is useful both for linguistic research and for adapting parsers to spoken English.

Key points

• The data covers sBNC, the spoken BNC section (about 10 million words, 1,035,527 sentences). The conversation part supplies 4 million of those words. Annotated progressives total 93,253: 78,580 have pronoun subjects, 1,509 have no subject, 83,305 are declarative and 9,948 interrogative.
• Auxiliary form splits into 38,015 full, 51,295 contracted and 3,943 zero. Tense splits into 70,938 present, 18,986 past, 2,956 present perfect and 373 past perfect.
• Register comes from the Davies (2006) genre labels, mapped onto three spontaneity levels: scripted (3,725), formal unscripted (39,101) and informal (42,165). Another 8,262 items are unclassified.
• A zero auxiliary appears in 82% of subjectless progressives and in 27% of second-person interrogatives (4,923 cases). Elsewhere the rate is only about 2%, and 4% across the whole set.
• Quality was checked by re-annotating a random 10% sample many months later, which gave more than 99% consistency. The release includes XML, a DTD and a README. The authors had earlier used statistics from these labels to improve probabilistic parsing of spoken data (Caines & Buttery, 2010).

Interesting ideas

• I1. The authors ship only labels plus document and sentence IDs, which lets licensed corpora be enriched and shared without redistributing their text.
• I2. Missing elements such as an absent subject or auxiliary get their own explicit "zero" value instead of being treated as noise. Ellipsis thus becomes a countable, analyzable category.
• I3. Clause type is decided purely by word order, kept apart from the pragmatic question of whether the speaker is asking something.
• I4. The authors fixed the grouping of genres into levels before any annotation started, which removed a source of judgment drift. Months later they re-labeled a random 10% to measure self-consistency.

For Mova

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

• I1 (learn): Silver labels and repairs for restricted treebanks or corpora could be published as stand-off files keyed by sentence ID, keeping the training data shareable.
• I2 (tree): The parser could mark elided auxiliaries and subjects explicitly in spoken-style input, so later stages know a constituent was omitted rather than misparsed.
• I3 (frames): The speech-act layer in prag could treat syntactic mood as a separate input from the act it infers, which helps with declarative questions like "You not having any cake?".
• I4 (learn): Category mappings for LLM-teacher or rule-based labeling could be frozen in advance, and a delayed random re-labeling run kept as a cheap stability check on the data.