基于SlovakBERT的命名实体识别

该模型是在Slovak wikiann数据集上对 gerulata/slovakbert 进行微调的版本。在评估集上取得以下结果：

• 损失：0.1600
• 精确度：0.9327
• 召回率：0.9470
• F1值：0.9398
• 准确度：0.9785

用途和限制

支持的类别：LOCATION、PERSON、ORGANIZATION

from transformers import pipeline


ner_pipeline = pipeline(task='ner', model='crabz/slovakbert-ner')
input_sentence = "Minister financií a líder mandátovo najsilnejšieho hnutia OĽaNO Igor Matovič upozorňuje, že následky tretej vlny budú na Slovensku veľmi veľké."
classifications = ner_pipeline(input_sentence)

使用displaCy的

import spacy
from spacy import displacy


ner_map = {0: '0', 1: 'B-OSOBA', 2: 'I-OSOBA', 3: 'B-ORGANIZÁCIA', 4: 'I-ORGANIZÁCIA', 5: 'B-LOKALITA', 6: 'I-LOKALITA'}

entities = []
for i in range(len(classifications)):
    if classifications[i]['entity'] != 0:
        if ner_map[classifications[i]['entity']][0] == 'B':
            j = i + 1
            while j < len(classifications) and ner_map[classifications[j]['entity']][0] == 'I':
                j += 1
            entities.append((ner_map[classifications[i]['entity']].split('-')[1], classifications[i]['start'],
                             classifications[j - 1]['end']))

nlp = spacy.blank("en")  # it should work with any language

doc = nlp(input_sentence)

ents = []
for ee in entities:
    ents.append(doc.char_span(ee[1], ee[2], ee[0]))

doc.ents = ents

options = {"ents": ["OSOBA", "ORGANIZÁCIA", "LOKALITA"],
           "colors": {"OSOBA": "lightblue", "ORGANIZÁCIA": "lightcoral", "LOKALITA": "lightgreen"}}
displacy_html = displacy.render(doc, style="ent", options=options)

如下：

import spacy
from spacy import displacy


ner_map = {0: '0', 1: 'B-OSOBA', 2: 'I-OSOBA', 3: 'B-ORGANIZÁCIA', 4: 'I-ORGANIZÁCIA', 5: 'B-LOKALITA', 6: 'I-LOKALITA'}

entities = []
for i in range(len(classifications)):
    if classifications[i]['entity'] != 0:
        if ner_map[classifications[i]['entity']][0] == 'B':
            j = i + 1
            while j < len(classifications) and ner_map[classifications[j]['entity']][0] == 'I':
                j += 1
            entities.append((ner_map[classifications[i]['entity']].split('-')[1], classifications[i]['start'],
                             classifications[j - 1]['end']))

nlp = spacy.blank("en")  # it should work with any language

doc = nlp(input_sentence)

ents = []
for ee in entities:
    ents.append(doc.char_span(ee[1], ee[2], ee[0]))

doc.ents = ents

options = {"ents": ["OSOBA", "ORGANIZÁCIA", "LOKALITA"],
           "colors": {"OSOBA": "lightblue", "ORGANIZÁCIA": "lightcoral", "LOKALITA": "lightgreen"}}
displacy_html = displacy.render(doc, style="ent", options=options)

Minister financií a líder mandátovo najsilnejšieho hnutia OĽaNO ORGANIZÁCIA Igor Matovič OSOBA upozorňuje, že následky tretej vlny budú na Slovensku LOKALITA veľmi veľké.

训练过程

训练超参数

训练过程中使用了以下超参数：

• 学习率：5e-05
• 训练批次大小：32
• 评估批次大小：8
• 种子：42
• 优化器：带有betas=(0.9,0.999)和epsilon=1e-08的Adam
• 学习率调度器类型：linear
• 训练轮数：15.0

训练结果

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
0.2342	1.0	625	0.1233	0.8891	0.9076	0.8982	0.9667
0.1114	2.0	1250	0.1079	0.9118	0.9269	0.9193	0.9725
0.0817	3.0	1875	0.1093	0.9173	0.9315	0.9243	0.9747
0.0438	4.0	2500	0.1076	0.9188	0.9353	0.9270	0.9743
0.028	5.0	3125	0.1230	0.9143	0.9387	0.9264	0.9744
0.0256	6.0	3750	0.1204	0.9246	0.9423	0.9334	0.9765
0.018	7.0	4375	0.1332	0.9292	0.9416	0.9353	0.9770
0.0107	8.0	5000	0.1339	0.9280	0.9427	0.9353	0.9769
0.0079	9.0	5625	0.1368	0.9326	0.9442	0.9383	0.9785
0.0065	10.0	6250	0.1490	0.9284	0.9445	0.9364	0.9772
0.0061	11.0	6875	0.1566	0.9328	0.9433	0.9380	0.9778
0.0031	12.0	7500	0.1555	0.9339	0.9473	0.9406	0.9787
0.0024	13.0	8125	0.1548	0.9349	0.9462	0.9405	0.9787
0.0015	14.0	8750	0.1562	0.9330	0.9469	0.9399	0.9788
0.0013	15.0	9375	0.1600	0.9327	0.9470	0.9398	0.9785

框架版本

• Transformers 4.13.0.dev0
• Pytorch 1.10.0+cu113
• Datasets 1.15.1
• Tokenizers 0.10.3