Named Entity Recognition for Chinese (BERT-Weibo Dataset)

Description

This model annotates named entities in a text, that can be used to find features such as names of people, places, and organizations. The model does not read words directly but instead reads word embeddings, which represent words as points such that more semantically similar words are closer together.

This model uses the pre-trained bert_base_chinese embeddings model from BertEmbeddings annotator as an input, so be sure to use the same embeddings in the pipeline.

Predicted Entities

Tag Meaning Example
PER.NAM Person name 张三
PER.NOM Code, category 穷人
LOC.NAM Specific location 紫玉山庄
LOC.NOM Generic location 大峡谷、宾馆
GPE.NAM Administrative regions and areas 北京
ORG.NAM Specific organization 通惠医院
ORG.NOM Generic or collective organization 文艺公司

Live Demo Open in Colab Download

How to use

...
word_segmenter = WordSegmenterModel.pretrained("wordseg_large", "zh")\
        .setInputCols(["sentence"])\
        .setOutputCol("token")
embeddings = BertEmbeddings.pretrained(name='bert_base_chinese', lang='zh')\
          .setInputCols("document", "token") \
          .setOutputCol("embeddings")
ner = NerDLModel.pretrained("ner_weibo_bert_768d", "zh") \
        .setInputCols(["document", "token", "embeddings"]) \
        .setOutputCol("ner")
...
pipeline = Pipeline(stages=[document_assembler, sentence_detector, word_segmenter, embeddings, ner, ner_converter])
example = spark.createDataFrame(pd.DataFrame({'text': ["""张三去中国山东省泰安市爬中国五岳的泰山了"""]}))
result = pipeline.fit(example).transform(example)
...
val word_segmenter = WordSegmenterModel.pretrained("wordseg_large", "zh")
     .setInputCols(Array("sentence"))
     .setOutputCol("token")
val embeddings = BertEmbeddings.pretrained(name='bert_base_chinese', lang='zh')
     .setInputCols(Array("document", "token"))
     .setOutputCol("embeddings")
val ner = NerDLModel.pretrained("ner_weibo_bert_768d", "zh")
     .setInputCols(Array("document", "token", "embeddings"))
     .setOutputCol("ner")
...
val pipeline = new Pipeline().setStages(Array(document_assembler, sentence_detector, word_segmenter, embeddings, ner))
val result = pipeline.fit(Seq.empty["张三去中国山东省泰安市爬中国五岳的泰山了"].toDS.toDF("text")).transform(data)
import nlu
text = ["张三去中国山东省泰安市爬中国五岳的泰山了"]

ner_df = nlu.load('zh.ner.weibo.bert_768d').predict(text)
ner_df

Results

+--------+-------+
|token   |ner    |
+--------+-------+
|张三    |PER.NAM|
|中国    |GPE.NAM|
|山东省  |GPE.NAM|
|中国五岳|GPE.NAM|
|泰山    |GPE.NAM|
+--------+-------+

Model Information

Model Name: ner_weibo_bert_768d
Type: ner
Compatibility: Spark NLP 2.7.0+
License: Open Source
Edition: Official
Input Labels: [sentence, token, embeddings]
Output Labels: [ner]
Language: zh

Data Source

The model was trained on the Weibo NER (He and Sun, 2017) data set.

Benchmarking

| ner_tag      | precision | recall | f1-score | support |
|--------------|-----------|--------|----------|---------|
| GPE.NAM      | 0.73      | 0.66   | 0.69     | 50      |
| GPE.NOM      | 0.00      | 0.00   | 0.00     | 2       |
| LOC.NAM      | 0.60      | 0.10   | 0.18     | 29      |
| LOC.NOM      | 0.20      | 0.10   | 0.13     | 10      |
| ORG.NAM      | 0.53      | 0.15   | 0.23     | 60      |
| ORG.NOM      | 0.50      | 0.28   | 0.36     | 18      |
| PER.NAM      | 0.66      | 0.61   | 0.63     | 139     |
| PER.NOM      | 0.68      | 0.63   | 0.66     | 197     |
| accuracy     | 0.96      | 9110   |          |         |
| macro avg    | 0.54      | 0.39   | 0.43     | 9110    |
| weighted avg | 0.96      | 0.96   | 0.96     | 9110    |