完善韩泰 EOT 标注编辑器并更新标注规范与项目资料

2026-10-09 15:55:46 +08:00
parent 138b08a8e4
commit 0204f1288a
29 changed files with 1340 additions and 1271 deletions
@@ -0,0 +1,143 @@
+"""Convert Hantai XLSX rows to an editable, approximate CSV draft.
+
+Requires ffmpeg and numpy. Never writes the source workbook or 工作数据.
+"""
+import csv
+import json
+import math
+import subprocess
+import uuid
+import zipfile
+from collections import defaultdict
+from pathlib import Path
+from xml.etree import ElementTree as ET
+
+import numpy as np
+
+ROOT = Path(__file__).resolve().parents[3]
+SOURCE = ROOT / "data/韩泰机器人抢话/韩泰机器人抢话测试数据集-已校对.xlsx"
+OUTPUT = ROOT / "labels/EOT标注/韩泰初稿"
+NS = {"s": "http://schemas.openxmlformats.org/spreadsheetml/2006/main"}
+
+
+def source_rows():
+    with zipfile.ZipFile(SOURCE) as archive:
+        strings = []
+        if "xl/sharedStrings.xml" in archive.namelist():
+            strings = ["".join(item.itertext()) for item in ET.fromstring(archive.read("xl/sharedStrings.xml"))]
+        rows = []
+        for row in ET.fromstring(archive.read("xl/worksheets/sheet1.xml")).findall(".//s:row", NS):
+            values = {}
+            for cell in row.findall("s:c", NS):
+                value = cell.find("s:v", NS)
+                text = value.text if value is not None else ""
+                if cell.get("t") == "s":
+                    text = strings[int(text)]
+                elif cell.get("t") == "inlineStr":
+                    text = "".join(cell.find("s:is", NS).itertext())
+                values[cell.get("r").rstrip("0123456789")] = text or ""
+            try:
+                uuid.UUID(values.get("A", ""))
+            except ValueError:
+                continue
+            rows.append({"row": int(row.get("r")), **values})
+        return rows
+
+
+def speech_spans(audio):
+    raw = subprocess.check_output(["ffmpeg", "-v", "error", "-i", str(audio), "-f", "f32le", "-ac", "2", "-ar", "8000", "pipe:1"])
+    samples = np.frombuffer(raw, dtype=np.float32).reshape(-1, 2)
+    frames = len(samples) // 160
+    rms = np.sqrt((samples[:frames * 160].reshape(frames, 160, 2) ** 2).mean(axis=1))
+    channels = []
+    for channel in range(2):
+        threshold = max(.004, float(np.quantile(rms[:, channel], .95)) * .16)
+        runs = []
+        for index in np.flatnonzero(rms[:, channel] > threshold):
+            if runs and index - runs[-1][1] < 35:
+                runs[-1][1] = int(index)
+            else:
+                runs.append([int(index), int(index)])
+        channels.append([(a * 20, (b + 1) * 20) for a, b in runs if b - a >= 12])
+    return channels, round(len(samples) / 8)
+
+
+def match_prompts(groups, spans):
+    # Greeting is absent from the source. Match remaining prompts in order,
+    # allowing extra robot speech (e.g. a retry) to be skipped.
+    candidates = spans[1:]
+    cache = {}
+
+    def solve(i, j):
+        if i == len(groups):
+            return .35 * (len(candidates) - j), []
+        if len(candidates) - j < len(groups) - i:
+            return float("inf"), []
+        if (i, j) in cache:
+            return cache[i, j]
+        expected = max(.6, len(groups[i][0]["C"]) / 6.7)
+        actual = (candidates[j][1] - candidates[j][0]) / 1000
+        score, rest = solve(i + 1, j + 1)
+        best = (score + abs(math.log(actual / expected)), [candidates[j]] + rest)
+        if len(candidates) - j > len(groups) - i:
+            score, rest = solve(i, j + 1)
+            if score + .35 < best[0]:
+                best = score + .35, rest
+        cache[i, j] = best
+        return best
+
+    matches = solve(0, 0)[1]
+    if len(matches) != len(groups):
+        raise ValueError("Not enough robot speech spans; inspect the recording before conversion.")
+    return matches
+
+
+def main():
+    grouped = defaultdict(list)
+    for row in source_rows():
+        grouped[row["A"]].append(row)
+    transcripts, annotations, provenance = [], [], []
+    for call, rows in grouped.items():
+        channels, duration = speech_spans(ROOT / f"data/韩泰机器人抢话/音频/{call}.mp3")
+        groups = []
+        for row in rows:
+            if groups and groups[-1][0]["C"] == row["C"]:
+                groups[-1].append(row)
+            else:
+                groups.append([row])
+        prompts = match_prompts(groups, channels[1])
+        for index, (group, (start, end)) in enumerate(zip(groups, prompts)):
+            transcripts.append([call, "agent", start, end, group[0]["C"]])
+            window_end = prompts[index + 1][0] if index + 1 < len(prompts) else duration
+            user_spans = [(a, b) for a, b in channels[0] if a >= end - 300 and a < window_end]
+            user_start = user_spans[0][0] if user_spans else min(end + 300, window_end - 1)
+            user_end = user_spans[-1][1] if user_spans else window_end
+            # Multiple answers to the same robot prompt share its response window.
+            # These boundaries are provisional, not word-level forced alignment.
+            weights = [max(1, len(row["B"])) for row in group]
+            cursor = user_start
+            for offset, row in enumerate(group):
+                finish = user_end if offset == len(group) - 1 else user_start + round((user_end - user_start) * sum(weights[:offset + 1]) / sum(weights))
+                finish = max(cursor + 1, min(duration, finish))
+                transcripts.append([call, "user", cursor, finish, row["B"]])
+                conflict = row.get("E") in ("是", "否") and row["E"] != row["D"]
+                label = "" if conflict else "0" if row["D"] == "是" else "1" if row["D"] == "否" else ""
+                reason = "uncertain" if label == "" else "unfinished_expression" if label == "0" else "completed_answer"
+                annotations.append([call, finish, row["B"], label, "low", reason, "xlsx_draft"])
+                provenance.append({"call_id": call, "source_row": row["row"], "decision_ms": finish, "original_interrupted": row["D"], "unheaded_column_e": row.get("E", ""), "conflict": conflict})
+                cursor = finish
+    OUTPUT.mkdir(parents=True, exist_ok=True)
+    for name, header, records in [
+        ("asr_history.csv", ["call_id", "role", "start_ms", "end_ms", "transcript"], sorted(transcripts, key=lambda r: (r[0], r[2], r[1]))),
+        ("eot_labels.csv", ["call_id", "decision_ms", "user_text_at_decision", "can_reply_now", "confidence", "reason", "annotator"], annotations),
+    ]:
+        with (OUTPUT / name).open("w", encoding="utf-8-sig", newline="") as file:
+            writer = csv.writer(file)
+            writer.writerow(header)
+            writer.writerows(records)
+    (OUTPUT / "source_rows.json").write_text(json.dumps(provenance, ensure_ascii=False, indent=2) + "\n")
+    print(json.dumps({"calls": len(grouped), "transcripts": len(transcripts), "annotations": len(annotations), "conflicts": sum(row["conflict"] for row in provenance)}))
+
+
+if __name__ == "__main__":
+    main()