VALIDATION.md

# Release validation

Generated by `build_hf_dataset.py` from the project's gold annotation sources. Rows were parsed with Python's `csv` module; quoted fields can contain commas, so shell line counts are not row counts.

| File | Rows | Columns | Empty fields | Duplicate rows | SHA-256 |
|---|---:|---:|---:|---:|---|
| `data/all.csv` | 878 | 14 | 1,971 | 0 | `9c8226ce8ab508db20918a086901b6521f9980a76ef75fd69fbd1b9e5127c08c` |
| `data/ted.csv` | 778 | 24 | 1,243 | 0 | `59b10f628a83b8237dff0e88d95bf8ebbbac9380be771e6f88a24d8b49383019` |
| `data/beat2.csv` | 100 | 25 | 97 | 0 | `80fa3543b4fe89ddfe0958a19c106fec03464c3a4a41ce9956a08e1eaf5fa96b` |
| `data/ted_talks.csv` | 10 | 9 | 0 | 0 | `dbfeaee26d652efa8188526f79374b7d2d90618952e6b41e893bdb29e4362fa3` |

`videos/beat2/` holds 99 MP4 files, one per distinct BEAT2 gesture (100 BEAT2 rows; one gesture is annotated twice). Empty fields are expected: TED-only and BEAT2-only columns are blank for the other source, and `reviewed_description` is blank outside the 231-sample reviewed subset.

## Paper-statistic checks

- Total clips: 878 (paper: 878) — matched.
- TED clips: 778 (paper: 778) from 10 talks (paper: 10) — matched.
- BEAT2 clips: 100 (paper: 100) — matched.
- TED reviewed subset: 231 (paper: 231).
- TED intent categories: 15 (paper: 15).

## Content checks

- TED `gold_description`, `communicative_intent`, `transcript`, `gesture_words` and `intent_category_raw` equal the gold spreadsheet after trimming surrounding whitespace.
- TED `physical_form` comes from the project's deterministic split (`ted_human_annotation_exact_split_v1`) and is verified to be a prefix of `gold_description` for every row.
- TED `reviewed_description` is the text used as the reference in the paper's anchor-quality evaluation on the 231-sample subset; it differs from `gold_description` in 80 rows.
- Each TED talk maps to TED's official YouTube upload; upload and source-video durations agree to within 0.11 s (per talk in `data/ted_talks.csv`), and every TED gesture ends before its upload does. `youtube_url` starts at `floor(start_sec)`; `youtube_embed_url` requests `floor(start_sec)` to `ceil(end_sec)`.
- BEAT2 transcripts match BEAT2 segment metadata; `start_frame`/`end_frame` equal `int(sec * 30)`.
- BEAT2 descriptions with surrounding whitespace or an unbalanced straight double quote removed (9 rows): `beat2_090`, `beat2_091`, `beat2_092`, `beat2_093`, `beat2_094`, `beat2_095`, `beat2_096`, `beat2_098`, `beat2_099`. No other text was edited.

## TED timeline alignment with YouTube

Checked with `check_yt_alignment.py` (kept with the build script): YouTube's seek-bar preview frames (160×90, one every 5–10 s over the whole talk) were compared with the source video, downscaled and sampled at 10 fps, for offsets from −3 s to +3 s in 0.1 s steps. Error is the mean absolute grey-level difference (0–255); blank preview frames are skipped.

| Talk | YouTube ID | Preview frames | Best offset (s) | Error at best | Error at 0 s | Error at −1 s / +1 s |
|---|---|---:|---:|---:|---:|---|
| t0001 | `0Fi83BHQsMA` | 175 | +0.0 | 4.76 | 4.76 | 15.58 / 15.26 |
| t0002 | `8jPQjjsBbIc` | 150 | +0.0 | 10.0 | 10.0 | 15.65 / 14.26 |
| t0003 | `lmyZMtPVodo` | 145 | +0.0 | 6.0 | 6.0 | 21.61 / 20.61 |
| t0004 | `MB5IX-np5fE` | 125 | +0.0 | 2.78 | 2.78 | 7.4 / 7.97 |
| t0005 | `1BXYSGepx7Q` | 120 | +0.0 | 6.89 | 6.89 | 10.94 / 11.63 |
| t0006 | `LeBKBFAPwNc` | 97 | -0.1 | 4.82 | 5.07 | 6.11 / 7.05 |
| t0007 | `BHY0FxzoKZE` | 158 | +0.0 | 4.16 | 4.16 | 15.1 / 13.6 |
| t0008 | `twG4mr6Jov0` | 169 | +0.0 | 3.8 | 3.8 | 10.75 / 9.56 |
| t0009 | `b5ZESpOAolU` | 164 | +0.0 | 5.75 | 5.75 | 18.06 / 17.0 |
| t0010 | `5J6jAC6XxAI` | 162 | +0.0 | 8.57 | 8.57 | 15.39 / 14.41 |

The sampled content supports a near-zero global offset for all ten talks. The check fits one offset per talk on a 0.1 s grid; it does not establish ±0.1 s accuracy for every gesture.

## Flags

| Flag | Source | Rows | Meaning |
|---|---|---:|---|
| `gesture_not_visible` | TED | 10 | Annotator noted the gesture is not visible in the video; the intent field holds that note instead of an intent. |
| `incomplete_physical_form` | TED | 7 | Physical form has 4 or fewer words (e.g. "Both hands make."). |
| `gesture_words_not_in_interval` | TED | 2 | The gesture words are spoken outside the annotated interval (found by the per-gesture link check below); the interval, transcript snippet and original clip agree with each other. |
| `shared_interval` | TED | 68 | Another row has exactly the same interval: the annotated region contains two gestures, told apart by `gesture_words`. |
| `duplicate_gesture` | BEAT2 | 2 | The same BEAT2 gesture was annotated twice with different descriptions; both are kept. |
| `no_clear_meaning` | BEAT2 | 1 | The annotator states the gesture has no discernible meaning. |

## Data notes

- `id` is the only unique key. `beat2_gesture_name` identifies a BEAT2 gesture and its video, but `beat2_007`, `beat2_096` share one gesture (annotated twice, both kept). `legacy_mapped_id` is an internal index kept for traceability and is not unique.
- TED physical-form text is shared within each gesture-lexicon entry: 342 entries cover the 778 TED clips.
- TED transcripts are Whisper ASR output; a few were corrected or trimmed by hand.
- BEAT2 frame indices are zero-based and end-exclusive at 30 fps (`int(sec * 30)`); BEAT2 videos cover the gesture with up to 5 s of context on each side, bounded by the recording.

## Per-gesture check of the TED YouTube links

`validate_ted_links.py` (kept with the build script) checked all 778 TED rows:

- **A. Annotated clip vs source video:** the original SEMANTIX clip matches the source talk video at the row's timestamps (best offset in a ±1 s sweep within ±0.2 s): 774/778 rows.
- **B. YouTube vs source near the gesture:** YouTube's preview frames within ±12 s of the gesture match the source video best at a local offset within ±0.5 s: 729/778 rows that have non-blank preview frames nearby.
- **C. Speech near the linked time (TED's transcript):** at least half of the gesture's content words appear in TED's own transcript within ±1.5 s of the gesture. TED.com cue times were shifted to the YouTube timeline by one offset per talk (11.0–11.8 s, the length of the TED intro on YouTube), cross-fitted so that each row is checked with an offset fitted on the other half of its talk: 775/778 rows.
- **D. Word timing (word-level ASR):** the source talks were transcribed with word timestamps (faster-whisper `small.en`, `asr_word_times.py`; no offset fitting). For every content word of the gesture words (stopwords ignored; numbers, hyphens and plurals normalised), the nearest ASR occurrence is located and its distance from the original annotated interval measured (0 if it overlaps). Runs of 3+ zero-duration ASR words, a Whisper repetition artefact, are ignored. The worst word decides the row's `link_check` value:

| `link_check` | Rows | Meaning |
|---|---:|---|
| `inside` | 751 | Every content word is spoken within 0.5 s of the interval |
| `near` | 17 | Every content word is spoken within 2 s of the interval |
| `far` | 5 | Some content word is more than 2 s away; the 2 rows more than 3 s away carry `gesture_words_not_in_interval` |
| `asr_unconfirmed` | 5 | ASR did not produce some content word (spelling or recognition variant); see the verdicts below |

These checks are automated evidence, not a manual audit of every gesture. For the rows below, the word timestamps are the nearest ASR occurrence of each content word (`word@start-end`, seconds).

| Row | `link_check` | Word timestamps | Verdict |
|---|---|---|---|
| `ted_0013` [67.03–71.94] | `near` | kidnapping@66.84-67.40 extortion@68.60-69.28 getting@70.18-70.42 drug@70.90-71.20 trade@71.20-71.58 illegal@72.14-72.50 mining@72.50-72.96 | All content words within 0.56 s of the interval. |
| `ted_0036` [213.55–217.03] | `near` | But@218.22-218.62 | All content words within 1.19 s of the interval. |
| `ted_0049` [282.48–284.58] | `asr_unconfirmed` | guerrilla: not found jungle@283.14-283.40 | Spelling only: ASR writes "guerillas" inside the interval. |
| `ted_0109` [746.90–748.00] | `far` | saving@755.64-755.96 place@756.16-756.36 | Substantial mismatch: the words are spoken about 8 s after the interval, while the speaker is off screen; the interval matches its transcript snippet and the original clip. Flagged. |
| `ted_0254` [553.46–555.28] | `near` | harmed@551.96-552.28 drug@555.12-555.38 helped@554.48-554.70 drug@555.12-555.38 | All content words within 1.18 s of the interval. |
| `ted_0257` [568.10–570.50] | `asr_unconfirmed` | begin@569.74-570.10 conversastion: not found | Typo in the gesture words ("conversastion"); ASR has "conversation" inside the interval. |
| `ted_0309` [133.94–134.75] | `asr_unconfirmed` | backward: not found | ASR produced no words in this 0.8 s interval; TED's transcript at the linked time contains them (C). |
| `ted_0388` [438.02–438.72] | `near` | policy@439.58-440.00 | All content words within 0.86 s of the interval. |
| `ted_0393` [459.49–461.75] | `near` | such@459.26-459.46 visceral@459.58-460.26 hatred@460.26-461.64 anger@462.98-463.38 | All content words within 1.23 s of the interval. |
| `ted_0432` [610.47–611.53] | `near` | we@612.16-612.28 would@612.28-612.42 not@612.42-612.60 | All content words within 0.89 s of the interval. |
| `ted_0451` [40.91–42.06] | `near` | understand@39.78-40.32 | All content words within 0.59 s of the interval. |
| `ted_0475` [126.24–127.47] | `near` | book@125.52-125.66 | All content words within 0.58 s of the interval. |
| `ted_0523` [304.69–305.51] | `near` | mean@303.48-303.68 | All content words within 1.01 s of the interval. |
| `ted_0558` [487.86–488.43] | `near` | goes@486.96-487.26 deeper@487.26-487.66 | All content words within 0.6 s of the interval. |
| `ted_0591` [661.39–662.77] | `near` | solve@661.82-662.16 problem@664.16-664.56 | All content words within 1.39 s of the interval. |
| `ted_0599` [700.87–702.72] | `near` | won't@701.54-701.70 lonely@701.90-702.12 won't@701.54-701.70 feel@702.58-702.78 like@702.78-702.94 life@702.94-703.08 pointless@703.28-703.60 | All content words within 0.56 s of the interval. |
| `ted_0602` [720.06–722.27] | `far` | didn't@717.50-717.66 know@717.66-717.78 anything@717.78-718.10 | Words end about 2 s before the interval starts; the interval is shared with `ted_0603`. |
| `ted_0733` [288.18–289.27] | `near` | And@287.10-287.26 I@287.26-287.98 | All content words within 0.92 s of the interval. |
| `ted_0735` [309.97–310.62] | `near` | going@309.02-309.12 explore@309.20-309.64 | All content words within 0.85 s of the interval. |
| `ted_0805` [245.36–246.63] | `near` | streaming@245.38-245.74 both@246.32-247.00 cheeks@247.26-247.56 | All content words within 0.63 s of the interval. |
| `ted_0878` [600.50–603.33] | `near` | means@599.04-599.74 citizen@599.74-600.94 negative@602.86-603.32 | All content words within 0.76 s of the interval. |
| `ted_0893` [662.35–664.65] | `asr_unconfirmed` | cherished: not found ideal@663.66-664.36 | ASR hears "cherish" for "cherished" inside the interval. |
| `ted_1160` [414.12–415.28] | `far` | moms@408.44-408.74 dads@409.12-409.44 | Substantial mismatch: the words are spoken about 5 s before the interval, which is shared with `ted_1161` (whose words do match). Flagged. |
| `ted_1215` [18.35–18.82] | `far` | first@21.04-21.22 impression@21.22-21.78 | Words start about 2 s after the interval ends; the interval is shared with `ted_1214`. |
| `ted_1285` [386.13–387.21] | `near` | eyes@387.88-388.24 | All content words within 0.67 s of the interval. |
| `ted_1322` [602.44–604.03] | `asr_unconfirmed` | getting@602.08-602.64 good: not found deal@603.02-603.30 | ASR hears "big deal" where TED's transcript at the linked time has "good deal" (C). |
| `ted_1338` [671.78–673.11] | `far` | just@669.28-669.56 Her@671.14-671.38 | "just" is spoken 2.2 s before the interval; this row also carries `gesture_not_visible`. |

17 further rows have their words inside the interval but were flagged by checks A–C:

| Row | Flags | Verdict |
|---|---|---|
| `ted_0020` | clip_offset=-0.5 | Words in the interval (D); the original clip file starts 0.5 s before the annotated time (A). |
| `ted_0115` | clip_offset=-0.7 | Words in the interval (D); the original clip file starts 0.7 s before the annotated time (A). |
| `ted_0142` | youtube_local_offset=+0.8 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_0234` | youtube_local_offset=+0.7 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_0280` | clip_offset=-0.4 | Words in the interval (D); the original clip file starts 0.4 s before the annotated time (A). |
| `ted_0714` | youtube_local_offset=+0.9 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_0715` | youtube_local_offset=+0.7 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_0762` | youtube_local_offset=-0.6 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_0808` | youtube_local_offset=+3.0 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_0834` | youtube_local_offset=-0.6 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_0858` | youtube_local_offset=-1.3 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_0877` | youtube_local_offset=-2.1 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_0888` | youtube_local_offset=+1.3 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_0921` | clip_offset=-0.3 | Words in the interval (D); the original clip file starts 0.3 s before the annotated time (A). |
| `ted_0943` | youtube_local_offset=-0.6 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |
| `ted_1137` | speech_recall=0.00 | Words in the interval (D); TED's transcript hyphenates "hyper-activated" (C). |
| `ted_1365` | youtube_local_offset=+0.9 | Words in the interval (D); nearby preview frames are a static shot, so B cannot pin down timing. |