new incoming dataset(s) of SWT transcription

[keighrim]

Issue for tracking progress of "transcribed-slates" and "transcribed-chyrons" sets re-uploaded to this repo.

[keighrim]

Just wanted to bring up that these two (and possibly more in line for the future) datasets contain manual annotations of

1. scene labeling (based on SWT categories)
2. verbatim text transcripts
3. "normalization" of personal names
4. structural understanding (or RFB)

And, at the moment in the current main branch (b035767)

.
├── batches
├── january-slates
├── newshour-chyron
├── newshour-namedentity
├── newshour-namedentity-wikipedialink
├── newshour-transcript-sync
├── role-filler-binding
├── role-filler-binding-seqtag
├── scene-recognition
├── ...

we actually have similar annotations in different annotation sub-projects. Concretely,

• january-slates
  • scene labeling (but with start/end marks, not point-wise)
    • not particular column for scene labels, every annotation is slate type
• newshour-chyron
  • scene labeling (but with start/end marks, not point-wise)
    • not particular column for scene labels, every annotation is chyron type
  • verbatim text transcripts
    • in text column
• role-filler-binding
  • scene labeling
    • in FRAME and SWT-TYPE columns
  • structural understanding
    • in ANNOTATIONS column
• role-filler-binding-seqtag
  • NO GOLD format decided
  • scene labeling
  • verbatim text transcripts (I assume these "cleaned text" is not 100% accurate though)
  • structural understanding
• scene-recognition
  • scene labeling
    • in timestamp and type-label columns

Each subproject has its own process.py to create gold files, but since there are "thematic" overlap between these annotation work, I think we might want to re-think the organization of this repo.

Another very relevant factor I want to mention is the difference between scene labeling annotation in january-slates/newshour-chyron work vs. scene-recognition/rfb*/(new)transcribed-* work.
To my understanding, the characteristics of these work should be mostly equivalent (granularity is different of course, so the quality of information is not completely "equal"), but the different ways the annotation is done (interval-wise, or point-wise) led us to issue like #86 and unusual implementation like #102 (subdirs under golds dir).

[keighrim]

we actually have similar annotations in different annotation sub-projects. Concretely,

• january-slates
  • scene labeling (but with start/end marks, not point-wise)
    • not particular column for scene labels, every annotation is slate type
• newshour-chyron
  • scene labeling (but with start/end marks, not point-wise)
    • not particular column for scene labels, every annotation is chyron type
  • verbatim text transcripts
    • in text column
• role-filler-binding
  • scene labeling
    • in FRAME and SWT-TYPE columns
  • structural understanding
    • in ANNOTATIONS column
• role-filler-binding-seqtag
  • NO GOLD format decided
  • scene labeling
  • verbatim text transcripts (I assume these "cleaned text" is not 100% accurate though)
  • structural understanding
• scene-recognition
  • scene labeling
    • in timestamp and type-label columns

I'd also like to take this as an opportunity to standardize column names used for the same-type annotations. Proposing

• timeframes: keep start / end
• timepoints: currently FRAME, timestamp --> at
• scene labels: currently SWT-TYPE, type-label --> scene-type
• text transcription: currently text --> text-transcript (as opposed to speech-transcript to reserve ASR-type annotation just in case)

[marcverhagen]

Sounds good to me. Definitely standardize {FRAME, timestamp} into one value, which could be at or even timestamp if the latter saves work; and standardize {SWT-TYPE, type-label} into scene-type or type-label or label (the latter might be better if we use this for various kinds of labels, but I do like the descriptiveness of scene-type).

I don't quite follow the motivation for changing text into text-transcript, but have no problem with it.