- Multiling 2019 维基百科词条提取任务
- 基于BERT，用CRF/NMT微调
- 论文见Multilingual Wikipedia Summarization and Title Generation On Low Resource Corpus
- 代码基于sberbank-ai的实现ner-bert
- Baseline1: Spacy多语言模型（七种）进行命名实体识别，直接挑选实体作为词条
- Baseline2: Spacy单语言模型（八种）进行依存剖析，提取主语作为词条
- Model1: BERT多语言模型+BiLSTM+Attn+CRF,对词条位置进行序列标注
- Model2: BERT多语言模型+BiLSTM+Attn+NMT,对词条位置进行序列标注(Seq2Seq)
- Model3: MUSE+BiLSTM+CRF，对词条位置进行序列标注
- 下载数据集，解压到
./WikiTrain19/full ./WikiTrain19/clipped - 运行
./preprocess.py，得到标注好的数据./WikiTrain19/tagged_data和纯语料./WikiTrain19/raw_data - 纯语料可以直接跑baseline
./simple_method.ipynb - 若要使用标注数据，运行
./combine_shuffle_divide.sh进行后处理，并下载BERT预训练多语言模型到./ner-bert/datadrive/multi_cased_L-12_H-768_A-12 - 运行
./ner-bert/examples/multiling-2019.ipynb ./ner-bert/examples/multiling-2019-nmt.ipynb进行BERT模型的训练，模型保存在./ner-bert/datadrive/models/multiling-2019/下
-
Baseline1
- en: 0.105
- de: 0.48
- es: 0.013
- fr: 0.397
- it: 0.021
- pt: 0.404
- ru: 0.25
-
Baseline2
- en: 0.507
- de: 0.490
- fr: 0.4
- es: 0.451
- el: 0.4
- pt: 0.433
- it: 0.523
- nl: 0.567
-
Model1（迭代十次）
precision recall f1-score support
B_MISC 0.665 0.443 0.532 1788
I_MISC 0.733 0.379 0.500 1348
micro avg 0.690 0.415 0.519 3136
macro avg 0.699 0.411 0.516 3136
weighted avg 0.694 0.415 0.518 3136
precision recall f1-score support
O 0.972 0.985 0.979 35451
MISC 0.601 0.447 0.513 1788
accuracy 0.959 37239
macro avg 0.787 0.716 0.746 37239
weighted avg 0.955 0.959 0.956 37239
- Model2（迭代十次）
precision recall f1-score support
B_MISC 0.890 0.910 0.900 1802
I_MISC 0.888 0.902 0.895 1339
micro avg 0.889 0.907 0.898 3141
macro avg 0.889 0.906 0.897 3141
weighted avg 0.889 0.907 0.898 3141
precision recall f1-score support
O 0.996 0.994 0.995 34664
MISC 0.888 0.916 0.902 1788
accuracy 0.990 36452
macro avg 0.942 0.955 0.948 36452
weighted avg 0.990 0.990 0.990 36452