Turkish Named Entity Recognition (NER) Model

这个模型是使用评估过的著名的土耳其NER数据集的经调优版本（ https://github.com/stefan-it/turkish-bert/files/4558187/nerdata.txt ）对"dbmdz/bert-base-turkish-cased"模型进行的微调。

微调参数：

task = "ner"
model_checkpoint = "dbmdz/bert-base-turkish-cased"
batch_size = 8 
label_list = ['O', 'B-PER', 'I-PER', 'B-ORG', 'I-ORG', 'B-LOC', 'I-LOC']
max_length = 512 
learning_rate = 2e-5 
num_train_epochs = 3 
weight_decay = 0.01 

如何使用：

model = AutoModelForTokenClassification.from_pretrained("akdeniz27/bert-base-turkish-cased-ner")
tokenizer = AutoTokenizer.from_pretrained("akdeniz27/bert-base-turkish-cased-ner")
ner = pipeline('ner', model=model, tokenizer=tokenizer, aggregation_strategy="first")
ner("your text here")

请参考" https://huggingface.co/transformers/_modules/transformers/pipelines/token_classification.html" "以了解带有聚合策略参数的实体分组。

参考测试结果：

• 准确率：0.9933935699477056
• F1值：0.9592969472710453
• 精确度：0.9543530277931161
• 召回率：0.9642923563325274

在 "Küçük, D., Küçük, D., Arıcı, N. 2016. Türkçe Varlık İsmi Tanıma için bir Veri Kümesi ("A Named Entity Recognition Dataset for Turkish"). IEEE Sinyal İşleme, İletişim ve Uygulamaları Kurultayı. Zonguldak, Türkiye." 论文中提到的测试集上进行的评估结果。

• 测试集 准确率 精确度 召回率 F1值
• 20010000 0.9946 0.9871 0.9463 0.9662
• 20020000 0.9928 0.9134 0.9206 0.9170
• 20030000 0.9942 0.9814 0.9186 0.9489
• 20040000 0.9943 0.9660 0.9522 0.9590
• 20050000 0.9971 0.9539 0.9932 0.9732
• 20060000 0.9993 0.9942 0.9942 0.9942
• 20070000 0.9970 0.9806 0.9439 0.9619
• 20080000 0.9988 0.9821 0.9649 0.9735
• 20090000 0.9977 0.9891 0.9479 0.9681
• 20100000 0.9961 0.9684 0.9293 0.9485
• 总体 0.9961 0.9720 0.9516 0.9617