icassp2027-experiment/docs/easycom_event_review_tool.md

EasyCom wearer-centric EOT/INT pre-label and review pipeline

The workflow has three explicit stages. Speech activity timestamps are read-only throughout; reviewed events never modify the dense 5 Hz activity ground truth.

EasyCom participant speech intervals
  -> floor-episode grouping + rule candidates + structured timing evidence
  -> DeepSeek semantic pre-label (optional but recommended)
  -> human Accept / Relabel / Exclude

1. Generate candidates

Develop the rules on training-side Sessions 4--12:

python3 prepare_easycom_eot_int_review.py \
  --dataset-root /path/to/easycom \
  --sessions Session_4 Session_5 Session_6 Session_7 Session_8 Session_9 Session_10 Session_11 Session_12 \
  --output-dir reports/easycom_floor_event_candidates

The generator emits two independent files:

• easycom_eot_int_candidates.jsonl contains floor-level wearer 0->1 behavior anchors retrospectively paired with a preceding non-wearer release (WEARER_ONSET, EOT_OPPORTUNITY_CANDIDATE), only episode-ending wearer 1->0 boundaries where a new non-wearer onset satisfies wearer_floor_start < other_onset < wearer_offset and the same activity crosses the offset (WEARER_OFFSET, YIELD_CANDIDATE), and hard-negative regions that need no wearer transition: SAME_SPEAKER_PAUSE, OTHER_TO_OTHER, OTHER_BACKCHANNEL, and NON_TAKEOVER_OVERLAP;
• easycom_eot_int_controls.jsonl contains short within-turn raw onsets (WEARER_RESUME), floor-level wearer entries with no surviving preceding release opportunity (WEARER_ONSET_CONTROL), and wearer offsets that fail the strict INT ordering/crossing prerequisite (WEARER_OFFSET_CONTROL). They are retained for audits, but are not sent to DeepSeek or placed in the primary human-review queue.

A raw wearer onset is merged into its preceding floor episode only when its preceding silence is at most 0.8 seconds and no plausible sustained non-wearer takeover occurs in that gap. Short feedback-like lexical or nonverbal activity is recorded but does not by itself split the episode. A non-feedback other contribution that occupies at least continuation_min_seconds of the gap or continues across the resumed onset conservatively preserves a new wearer-entry candidate. Every decision is stored in temporal_evidence, including the raw gap, feedback IDs, competing IDs, threshold, and floor_episode_merged flag.

For each floor-level wearer onset, EOT construction searches backward for the most recent non-wearer release and rejects the pair if wearer speech or a competing non-wearer onset intervenes. Silence length alone does not invalidate the opportunity. Each primary EOT candidate stores both opportunity_anchor_seconds and behavior_anchor_seconds, plus their response delay.

Paired wearer onsets and strictly qualified offsets are initially unsure; timing alone is not a semantic positive label. The INT prerequisite is wearer_floor_start < other.start < wearer_offset < other.end, with a configurable one-frame (50 ms) tolerance only at the offset boundary. Pre-existing other activity and raw offsets inside a merged wearer episode stay in the controls file. Hard-negative scans are rule hints and still require AI and human review. The JSONL retains candidate source, original event time, participant-level activity, raw durations/gaps/overlap evidence, transcript, and Target_of_Speech when present. Target_of_Speech is annotation assistance only and cannot be a deployable model input.

All heuristic durations are configurable rather than frozen paper rules:

--relation-window-seconds
--same-speaker-pause-max-seconds
--other-handover-max-seconds
--short-vocalization-max-seconds
--continuation-min-seconds
--int-offset-tolerance-seconds
--wearer-resume-merge-max-seconds

Use --max-candidates 1 for a one-event smoke test. summary.json records the exact configuration and candidate counts. easycom_session_timelines.json stores the read-only participant speech lanes used by the full-session web timeline; it is not sent to DeepSeek.

2. Add DeepSeek pre-labels

Prompt v6 treats the EasyCom transcript marker [L] as laughter-only evidence, not as a lexical floor take. Wearer laughter during another participant's retained turn is directed toward EOT_WEARER_BACKCHANNEL; non-wearer laughter overlapping a retained wearer turn is directed toward INT_OTHER_BACKCHANNEL. Ambiguous laughter remains UNSURE, and [L] alone is neither an exclusion reason nor evidence of a visible smile. The prompt also treats the two EOT anchors and strict INT ordering as deterministic prerequisites rather than asking the model to reconstruct them.

needs_human_review is not requested from DeepSeek: every pre-label enters the human queue by protocol, so the adjudicator assigns that boolean locally. This prevents a string-valued model field from invalidating and retrying an otherwise valid batch.

First inspect exactly what would be uploaded:

python3 adjudicate_easycom_eot_int_candidates.py \
  --input reports/easycom_floor_event_candidates/easycom_eot_int_candidates.jsonl \
  --output reports/easycom_floor_event_candidates/easycom_eot_int_deepseek.jsonl \
  --model deepseek-v4-flash \
  --max-events 1 \
  --dry-run

After setting DEEPSEEK_API_KEY and obtaining authorization to upload the candidate excerpts:

python3 adjudicate_easycom_eot_int_candidates.py \
  --input reports/easycom_floor_event_candidates/easycom_eot_int_candidates.jsonl \
  --output reports/easycom_floor_event_candidates/easycom_eot_int_deepseek.jsonl \
  --model deepseek-v4-flash \
  --allow-external-transcript-upload

DeepSeek sees candidate source, structured timing, per-speaker activity, and transcript together. It returns one of POSITIVE, HARD_NEGATIVE, NOT_APPLICABLE, EXCLUDE, or UNSURE; positives additionally receive EOT_RESPONSE or INT_YIELD, and hard negatives receive a named subtype. It cannot modify timestamps. The output preserves both the original rule candidate and the complete AI decision/provenance.

During live adjudication, the CLI displays a dependency-free progress bar over all candidate/pass work, including the current pass, batch, percentage, and elapsed time. Redirected logs emit periodic progress lines instead of carriage return animation.

As defense in depth, the adjudicator also detects and skips non-crossing WEARER_OFFSET rows from older v2 candidate files before making any API call.

3. Human review

Run the web workspace on the AI-enriched JSONL:

python3 review_floor_events.py \
  --dataset-root /path/to/easycom \
  --events reports/easycom_floor_event_candidates/easycom_eot_int_deepseek.jsonl \
  --output reports/easycom_event_reviews/reviews.jsonl

Open http://127.0.0.1:8765. The page shows a session candidate overview, participant activity near the event, video/audio, transcript, structured timing, and the AI verdict/confidence/reason. A decision guide first states the behavioral question, then presents the exact AI verdict/subtype with a plain-language meaning; the Accept control repeats that exact label so its effect is explicit. The common path is one operation:

• Accept keeps the AI verdict/type;
• Relabel expands verdict and subtype controls;
• Exclude requires a concrete data-quality reason.

The current primary-review UI deliberately has no second-review action. Older saved second_review records remain readable for backward compatibility.

Boundary correction is hidden under an optional disclosure and is limited to one 200 ms bin. The original timestamp is always retained. Each save atomically rewrites JSONL and CSV while preserving the rule candidate, AI decision, human decision, relabel flag, exclusion state, confidence, and provenance. The current UI does not ask for a reviewer name; the export retains an empty reviewer field only for compatibility with older review files.

The primary review queue also filters legacy non-crossing WEARER_OFFSET rows and explicit WEARER_RESUME controls. To audit the controls intentionally, pass --events .../easycom_eot_int_controls.jsonl --include-controls and use a separate review output file. The old --include-offset-controls spelling remains an alias.

Protocol wording for the paper appendix:

Raw wearer speech onsets are first grouped into floor-level speech episodes by merging short within-turn resumptions that are not separated by a competing floor takeover. Only the resulting floor-level 0→1 transitions are proposed as EOT-response candidates.

The original speech intervals are preserved unchanged for dense forecasting and auditability; this grouping is used only for behavioral event annotation.

Freeze discipline

Develop the guide and resolve disagreements only on Sessions 4--12. Freeze categories and thresholds before reviewing official test Sessions 1--3. Use two independent human review files plus a separate adjudication output for paper- quality labels. DeepSeek is pre-label metadata, not an independent reviewer.