home edit page issue tracker

This page pertains to UD version 2.

It appears that you have Javascript disabled. Please consider enabling Javascript for this page to see the visualizations.

Metadata harmonisation: align spoken-data fields with UniDive naming conventions

Back to English GUM · Back to index

Repo: https://github.com/UniversalDependencies/UD_English-GUM

Cross-posting from the UniDive WG1 T1.5 (spoken language guidelines) metadata harmonisation review. We compared UD_English-GUM’s current CoNLL-U metadata against the proposed naming conventions (see also the full treebank status table). This is a suggestion for maintainers to review - feel free to push back on anything that doesn’t fit the corpus. The comparison was carried out semi-automatically with the help of Claude (Anthropic); errors or misunderstandings are possible, so please double-check anything unclear.

1. Mark spoken documents with `# modality`

The spoken portion of this treebank is identifiable via meta::genre. Checked directly against en_gum-ud-dev.conllu and en_gum-ud-test.conllu (30 documents/1575 sentences in dev, 30 documents/1464 sentences in test; 15 genres, 2 documents each per split). 5 of the 15 genres are spoken - vlog, speech, podcast, court, conversation - 10 documents per split, 20 total:

Genre	dev	test
`conversation`	`GUM_conversation_grounded`, `GUM_conversation_risk`	`GUM_conversation_lambada`, `GUM_conversation_retirement`
`court`	`GUM_court_loan`, `GUM_court_negligence`	`GUM_court_insanity`, `GUM_court_mitigation`
`podcast`	`GUM_podcast_bangladesh`, `GUM_podcast_wrestling`	`GUM_podcast_bezos`, `GUM_podcast_multitasking`
`speech`	`GUM_speech_impeachment`, `GUM_speech_inauguration`	`GUM_speech_austria`, `GUM_speech_newzealand`
`vlog`	`GUM_vlog_portland`, `GUM_vlog_radiology`	`GUM_vlog_london`, `GUM_vlog_studying`

The other 10 genres (academic, bio, essay, fiction, interview, letter, news, textbook, voyage, whow) are written or mixed-modality (interview here is written Q&A, not transcribed speech).

Suggestion: Add # modality = spoken to the 20 documents listed above and # modality = written to the rest. This was checked in dev/test only - please confirm the same genre labels apply consistently in train.

We also checked source/summary metadata for one document per spoken genre to propose interaction parameters (an optional add-on) - please confirm per-document, since genre alone doesn’t guarantee every document in a genre matches exactly:

Genre	`degree_of_spontaneity`	`number_of_participants`	`context`	`setting`	`channels`	`symmetry`
`conversation`	unplanned	dialogic/multi-party	private	face-to-face	phonic-auditory; gestural-visual	symmetric
`court`	planned	multi-party	professional	face-to-face	phonic-auditory; gestural-visual	asymmetric
`podcast`	unplanned	multi-party	public	online	phonic-auditory	asymmetric
`speech`	planned	monologic	public	broadcast	phonic-auditory; gestural-visual	asymmetric
`vlog`	planned	monologic	public	online	phonic-auditory; gestural-visual	asymmetric

Notes: conversation is Santa Barbara Corpus audio (no video) of private family/partner talk; court is audio-only Supreme Court oral argument (live Q&A elicited by justices’ questions, hence elicited rather than unplanned); podcast (Global Voices) has editors plus 2 guests, audio-only; speech and vlog are both video-recorded (C-SPAN and YouTube respectively) with a single speaker/creator addressing an audience with no back-and-forth, hence asymmetric.

2. Document-level (naming conventions)

Field	Suggestion
`document_id`	make tags: doc_id
`meta::genre`	make tags: genre

3. Speaker-level (naming conventions)

Field	Suggestion
`speaker`	make tags: speaker_id

4. Sentence-level (naming conventions)

Field	Suggestion
`meta`	corpus-specific (sentence-level) - verify against metadata.html
`trailing_xml`	corpus-specific (sentence-level) - verify against metadata.html

Implementation notes

Quick search & replace: document_id→tag doc_id, meta::genre→tag genre, speaker→tag speaker_id: python3 workgroups/spoken-data/scripts/harmonize_metadata.py rename-comment DIR --map meta::genre=genre,speaker=speaker_id --write (the doc_id tag addition is a tagset change, not a comment rename).
Needs a small script: modality tagging by genre is mechanical and verified against the real corpus: python3 workgroups/spoken-data/scripts/harmonize_metadata.py tag-modality DIR --spoken-if '_(vlog|speech|podcast|court|conversation)_' --written-if '.*' --write gives exactly 10/30 spoken documents in dev and 10/30 in test, matching the draft’s count precisely. Only run on en_gum-ud-dev.conllu/en_gum-ud-test.conllu for now (train genre labels aren’t yet confirmed - see manual item below); once confirmed, the same command run over the whole treebank directory covers train too.
Needs manual input from maintainers:
- Confirm the 15 genre labels (and the 5 spoken ones) apply consistently in en_gum-ud-train.conllu before running the modality-tagging command over train.
- The proposed per-genre interaction parameters (degree_of_spontaneity, number_of_participants, context, setting, channels, symmetry) are genre-level defaults, not verified per document - once confirmed, adding them is a mechanical rename-comment-style insert per document, but the values themselves need sign-off first.
- meta and trailing_xml (sentence-level) are corpus-specific with no proposed names yet.