#!/usr/bin/env python3 """Restore the transcripts MR-Bench-H does not redistribute, from your own copies of the source datasets. Writes a copy of the items file in which every "transcript": null is filled (and, for IEMOCAP, "source_label" = (valence - 3) / 2 from EmoEvaluation). Texts are produced exactly as in the reference pipeline: CMU-MOSI / CMU-MOSEI / CH-SIMS / CH-SIMS v2.0 -- label.csv or meta.csv of the raw release (columns video_id, clip_id, text), or an MMSA processed pickle (aligned_50.pkl / unaligned_39.pkl / sims_unaligned.pkl: raw_text + id) MUStARD++ --mustardpp mustard++_text.csv (columns KEY, SENTENCE) MELD --meld folder with train_sent_emo.csv, dev_sent_emo.csv, test_sent_emo.csv (curly quotes -> straight) IEMOCAP --iemocap IEMOCAP_full_release (Session*/dialog/transcriptions + EmoEvaluation; bracketed non-verbal annotations such as [LAUGHTER] are removed, as they would leak the other channels) python tools/fill_transcripts.py --items data/test.jsonl --out data/test.local.jsonl \ --iemocap /data/IEMOCAP_full_release --sims /data/SIMS/label.csv --simsv2 /data/chsimsv2/meta.csv \ --mustardpp /data/mustardpp/mustard++_text.csv python tools/fill_transcripts.py --items data/test.jsonl --out /dev/null --mosi /data/MOSI/label.csv --verify Do not redistribute the filled file: it contains text whose licence does not allow it (IEMOCAP in particular). """ import argparse import csv import glob import json import os import pickle import re import sys MMSA = ("mosi", "mosei", "sims", "simsv2") def mmsa_texts(path): """{(split or None, id): text} from a label/meta CSV or an MMSA pickle.""" out = {} if path.endswith(".pkl"): with open(path, "rb") as f: blob = pickle.load(f) for split, d in blob.items(): if isinstance(d, dict) and "raw_text" in d and "id" in d: for i, t in zip(d["id"], d["raw_text"]): out[str(i)] = str(t) return out with open(path, encoding="utf-8") as f: for r in csv.DictReader(f): out[f"{r['video_id']}$_${r['clip_id']}"] = r["text"] return out def mustardpp_texts(path): with open(path, encoding="utf-8", errors="replace") as f: return {r["KEY"]: r["SENTENCE"] for r in csv.DictReader(f)} def meld_texts(folder): out = {} for split in ["train", "dev", "test"]: p = os.path.join(folder, f"{split}_sent_emo.csv") if not os.path.exists(p): continue with open(p, encoding="utf-8", errors="replace") as f: for r in csv.DictReader(f): t = r["Utterance"].replace("’", "'").replace("‘", "'").replace("“", '"').replace("”", '"') out[(split, f"dia{r['Dialogue_ID']}_utt{r['Utterance_ID']}")] = t return out def iemocap_texts(root): """{utterance id: (text, source_label)}""" hits = glob.glob(os.path.join(root, "Session1")) + glob.glob(os.path.join(root, "*", "Session1")) if not hits: sys.exit(f"--iemocap: no Session1 folder under {root}") top = os.path.dirname(hits[0]) out = {} for sess in range(1, 6): sd = os.path.join(top, f"Session{sess}", "dialog") trans, val = {}, {} for f in glob.glob(os.path.join(sd, "transcriptions", "*.txt")): for line in open(f, encoding="utf-8", errors="replace"): m = re.match(r"(Ses\S+) \[(\d+\.\d+)-(\d+\.\d+)\]: (.*)", line.strip()) if m: trans[m.group(1)] = re.sub(r"\s+", " ", re.sub(r"\[[A-Za-z_ ]+\]", " ", m.group(4))).strip() for f in glob.glob(os.path.join(sd, "EmoEvaluation", "*.txt")): for line in open(f, encoding="utf-8", errors="replace"): m = re.match(r"\[(\d+\.\d+) - (\d+\.\d+)\]\s+(\S+)\s+(\S+)\s+\[(\d+\.\d+), (\d+\.\d+), (\d+\.\d+)\]", line.strip()) if m: val[m.group(3)] = float(m.group(5)) for u, t in trans.items(): out[u] = (t, round((val[u] - 3.0) / 2.0, 3) if u in val else None) return out def main(): ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) ap.add_argument("--items", required=True) ap.add_argument("--out", required=True) for s in MMSA: ap.add_argument(f"--{s}", default=None, help="label.csv / meta.csv or MMSA .pkl") ap.add_argument("--mustardpp", default=None, help="mustard++_text.csv") ap.add_argument("--meld", default=None, help="folder with *_sent_emo.csv") ap.add_argument("--iemocap", default=None, help="IEMOCAP_full_release") ap.add_argument("--overwrite", action="store_true", help="also replace transcripts that are already present") ap.add_argument("--verify", action="store_true", help="compare present transcripts with your copy and report") args = ap.parse_args() lookup = {} for s in MMSA: if getattr(args, s): t = mmsa_texts(getattr(args, s)) lookup[s] = lambda it, t=t: t.get(it["source_id"]) if args.mustardpp: t = mustardpp_texts(args.mustardpp) lookup["mustardpp"] = lambda it, t=t: t.get(it["source_id"]) if args.meld: t = meld_texts(args.meld) lookup["meld"] = lambda it, t=t: t.get((it["source_split"], it["source_id"])) iemo = iemocap_texts(args.iemocap) if args.iemocap else None if iemo is not None: lookup["iemocap"] = lambda it: (iemo.get(it["source_id"]) or (None, None))[0] stats = {} with open(args.items, encoding="utf-8") as f, open(args.out, "w", encoding="utf-8") as g: for line in f: it = json.loads(line) s = it["source"] st = stats.setdefault(s, {"items": 0, "filled": 0, "not_found": 0, "still_missing": 0, "verified": 0, "differs": 0}) st["items"] += 1 if s in lookup: txt = lookup[s](it) if it.get("transcript") is not None and args.verify and txt is not None: st["verified" if txt == it["transcript"] else "differs"] += 1 if it.get("transcript") is None or args.overwrite: if txt is None: st["not_found"] += 1 else: it["transcript"], it["transcript_status"] = txt, "filled_locally" st["filled"] += 1 if s == "iemocap" and it.get("source_label") is None: it["source_label"] = (iemo.get(it["source_id"]) or (None, None))[1] st["still_missing"] += it.get("transcript") is None g.write(json.dumps(it, ensure_ascii=False) + "\n") print("| source | items | filled | not found | still missing | verified equal | differs |\n|---|---|---|---|---|---|---|") for s, st in stats.items(): print(f"| {s} | {st['items']} | {st['filled']} | {st['not_found']} | {st['still_missing']} | {st['verified']} | {st['differs']} |") print(f"-> {args.out} (local use only: do not redistribute)") if __name__ == "__main__": main()