Named Entity Recognition for Bengali (GloVe 840B 300d)

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 glove_840B_300 embeddings model from WordEmbeddings annotator as an input, so be sure to use the same embeddings in the pipeline.

Predicted Entities

Persons - PER, Locations - LOC, Organizations - ORG, Object - OBJ, and Others - O.

Live Demo Open in Colab Download

How to use

...
embeddings = WordEmbeddingsModel.pretrained("glove_840B_300", "xx")\
          .setInputCols("document", "token") \
          .setOutputCol("embeddings")
ner = NerDLModel.pretrained("ner_jifs_glove_840B_300d", "bn") \
        .setInputCols(["document", "token", "embeddings"]) \
        .setOutputCol("ner")
pipeline = Pipeline(stages=[document_assembler, sentence_detector, tokenizer, embeddings, ner])
example = spark.createDataFrame(pd.DataFrame({'text': ["৯০ এর দশকের শুরুর দিকে বৃহৎ আকারে মার্কিন যুক্তরাষ্ট্রে এর প্রয়োগের প্রক্রিয়া শুরু হয়'"]}))
result = pipeline.fit(example).transform(example)
...
val embeddings = WordEmbeddingsModel.pretrained("glove_840B_300", "xx")
        .setInputCols(Array("document", "token"))
        .setOutputCol("embeddings")
val ner = NerDLModel.pretrained("ner_jifs_glove_840B_300d", "bn")
        .setInputCols(Array("document", "token", "embeddings"))
        .setOutputCol("ner")
val pipeline = new Pipeline().setStages(Array(document_assembler, sentence_detector, tokenizer, embeddings, ner))
val result = pipeline.fit(Seq.empty["৯০ এর দশক  ি হৎ আক ি তর এর রয়ো রকি়া  হয়"].toDS.toDF("text")).transform(data)
import nlu

text = ["৯০ এর দশকের শুরুর দিকে বৃহৎ আকারে মার্কিন যুক্তরাষ্ট্রে এর প্রয়োগের প্রক্রিয়া শুরু হয়"]
ner_df = nlu.load('bn.ner').predict(text, output_level='token')
ner_df

Results

+-------------+-----+
|token        |ner  |
+-------------+-----+
|৯০           |O    |
|এর           |O    |
|দশকের        |O    |
|শুরুর        |O    |
|দিকে         |O    |
|বৃহৎ         |O    |
|আকারে        |O    |
|মার্কিন      |B-LOC|
|যুক্তরাষ্ট্রে|I-LOC|
|এর           |O    |
|প্রয়োগের    |O    |
|প্রক্রিয়া   |O    |
|শুরু         |O    |
|হয়          |O    |
|'            |O    |
+-------------+-----+

Model Information

Model Name: ner_jifs_glove_840B_300d
Type: ner
Compatibility: Spark NLP 2.7.0+
License: Open Source
Edition: Official
Input Labels: [sentence, token, embeddings]
Output Labels: [ner]
Language: bn

Data Source

The model was trained on the Bengali NER data set introduced in the Journal of Intelligent & Fuzzy Systems.

Reference:

  • Karim, Redwanul & Islam, M. A. & Simanto, Sazid & Chowdhury, Saif & Roy, Kalyan & Neon, Adnan & Hasan, Md & Firoze, Adnan & Rahman, Mohammad. (2019). A step towards information extraction: Named entity recognition in Bangla using deep learning. Journal of Intelligent & Fuzzy Systems. 37. 1-13. 10.3233/JIFS-179349.

Benchmarking

|              | precision | recall | f1-score | support |
|--------------|-----------|--------|----------|---------|
| B-LOC        | 0.81      | 0.72   | 0.76     | 2005    |
| B-OBJ        | 0.66      | 0.08   | 0.13     | 573     |
| B-ORG        | 0.67      | 0.31   | 0.42     | 853     |
| B-PER        | 0.76      | 0.76   | 0.76     | 4035    |
| I-LOC        | 0.64      | 0.52   | 0.58     | 357     |
| I-OBJ        | 0.00      | 0.00   | 0.00     | 57      |
| I-ORG        | 0.65      | 0.37   | 0.47     | 516     |
| I-PER        | 0.76      | 0.73   | 0.74     | 1223    |
| O            | 0.93      | 0.97   | 0.95     | 39499   |
| accuracy     |           |        | 0.90     | 49118   |
| macro avg    | 0.65      | 0.49   | 0.54     | 49118   |
| weighted avg | 0.89      | 0.90   | 0.89     | 49118   |