Qwen3ForTokenClassificationによる英語Universal Dependenciesの品詞付与

2024年12月19日の記事の手法をQwen3-0.6Bに適用して、UD_English-EWTのUPOS品詞付与を試してみた。Google Colaboratory (GPU版)だと、こんな感じ。

!pip install transformers datasets evaluate seqeval accelerate
s='$1=="transformers"{printf("-b v%s",$2)}'
!test -d transformers || git clone `pip3 list | awk '{s}'` https://github.com/huggingface/transformers
!test -d UD_English-EWT || git clone --depth=1 https://github.com/UniversalDependencies/UD_English-EWT
import json
def makejson(conllu_file,json_file):
  with open(conllu_file,"r",encoding="utf-8") as r, open(json_file,"w",encoding="utf-8") as w:
    d={"tokens":[],"tags":[]}
    for s in r:
      if s.strip()=="":
        if d["tokens"]>[]:
          print(json.dumps(d),file=w)
        d={"tokens":[],"tags":[]}
      else:
        t=s.split("\t")
        if len(t)==10 and t[0].isdecimal():
          d["tokens"].append(t[1])
          d["tags"].append(t[3])
makejson("UD_English-EWT/en_ewt-ud-train.conllu","train.json")
makejson("UD_English-EWT/en_ewt-ud-dev.conllu","dev.json")
makejson("UD_English-EWT/en_ewt-ud-test.conllu","test.json")
!env WANDB_DISABLED=true python3 transformers/examples/pytorch/token-classification/run_ner.py --task_name pos --model_name_or_path Qwen/Qwen3-0.6B --train_file train.json --validation_file dev.json --test_file test.json --output_dir ./Qwen3-0.6B-english-upos --overwrite_output_dir --do_train --do_eval --do_predict

私(安岡孝一)の手元では、1時間40分ほどで以下のmetricsが出力されて、Qwen3-0.6B-english-uposが出来上がった。

***** train metrics *****
  epoch                    =        3.0
  total_flos               =  4296208GF
  train_loss               =     0.1966
  train_runtime            = 1:28:11.70
  train_samples            =      12544
  train_samples_per_second =      7.112
  train_steps_per_second   =      0.889

***** eval metrics *****
  epoch                   =        3.0
  eval_accuracy           =     0.9154
  eval_f1                 =     0.8995
  eval_loss               =     0.3669
  eval_precision          =     0.9006
  eval_recall             =     0.8983
  eval_runtime            = 0:00:28.48
  eval_samples            =       2001
  eval_samples_per_second =     70.243
  eval_steps_per_second   =      8.811

***** predict metrics *****
  predict_accuracy           =     0.9169
  predict_f1                 =     0.9003
  predict_loss               =     0.3495
  predict_precision          =     0.9007
  predict_recall             =     0.8999
  predict_runtime            = 0:00:29.87
  predict_samples_per_second =      69.53
  predict_steps_per_second   =      8.704

eval・predictともにF1値が0.90程度で、正直あまり良くない。ちょっと動かしてみよう。

from transformers import pipeline
nlp=pipeline("token-classification","Qwen3-0.6B-english-upos")
print(nlp("It don't mean a thing if it ain't got that swing"))

出来立てのQwen3-0.6B-english-uposで「It don't mean a thing if it ain't got that swing」に品詞付与してみたところ、私の手元では以下の結果が得られた。

[{'entity': 'PRON', 'score': np.float32(0.9998696), 'index': 0, 'word': 'It', 'start': 0, 'end': 2}, {'entity': 'AUX', 'score': np.float32(0.8005378), 'index': 1, 'word': 'Ġdon', 'start': 2, 'end': 6}, {'entity': 'PART', 'score': np.float32(0.99996376), 'index': 2, 'word': "'t", 'start': 6, 'end': 8}, {'entity': 'VERB', 'score': np.float32(0.99999404), 'index': 3, 'word': 'Ġmean', 'start': 8, 'end': 13}, {'entity': 'DET', 'score': np.float32(0.99962735), 'index': 4, 'word': 'Ġa', 'start': 13, 'end': 15}, {'entity': 'NOUN', 'score': np.float32(0.9999932), 'index': 5, 'word': 'Ġthing', 'start': 15, 'end': 21}, {'entity': 'SCONJ', 'score': np.float32(1.0), 'index': 6, 'word': 'Ġif', 'start': 21, 'end': 24}, {'entity': 'PRON', 'score': np.float32(0.9999958), 'index': 7, 'word': 'Ġit', 'start': 24, 'end': 27}, {'entity': 'AUX', 'score': np.float32(0.9502168), 'index': 8, 'word': 'Ġain', 'start': 27, 'end': 31}, {'entity': 'PART', 'score': np.float32(0.9999628), 'index': 9, 'word': "'t", 'start': 31, 'end': 33}, {'entity': 'VERB', 'score': np.float32(0.9777384), 'index': 10, 'word': 'Ġgot', 'start': 33, 'end': 37}, {'entity': 'ADV', 'score': np.float32(0.49614692), 'index': 11, 'word': 'Ġthat', 'start': 37, 'end': 42}, {'entity': 'NOUN', 'score': np.float32(0.9643484), 'index': 12, 'word': 'Ġswing', 'start': 42, 'end': 48}]

空白が「Ġ」に化けてしまっているのは、まあ何とかするとしても、「that」が読めていない。やはりQwen3ForTokenClassificationも、品詞付与の精度がイマイチ上がらないようだ。うーん、残念。