Korean Lemmatizer


This model uses context and language knowledge to assign all forms and inflections of a word to a single root. This enables the pipeline to treat the past and present tense of a verb, for example, as the same word instead of two completely different words. The lemmatizer takes into consideration the context surrounding a word to determine which root is correct when the word form alone is ambiguous.

Live Demo Open in Colab Download

How to use

document_assembler = DocumentAssembler() \
    .setInputCol("text") \
word_segmenter = WordSegmenterModel.pretrained('wordseg_kaist_ud', 'ko')\
lemmatizer = LemmatizerModel.pretrained("lemma", "ko") \
        .setInputCols(["token"]) \
nlp_pipeline = Pipeline(stages=[document_assembler, word_segmenter , lemmatizer])
light_pipeline = LightPipeline(nlp_pipeline.fit(spark.createDataFrame([[""]]).toDF("text")))
results = light_pipeline.fullAnnotate(["이렇게되면이러한인간형을다투어본받으려할것이틀림없다."])
val document_assembler = DocumentAssembler()
val word_segmenter = WordSegmenterModel.pretrained("wordseg_kaist_ud", "ko")
val lemmatizer = LemmatizerModel.pretrained("lemma", "ko")
val pipeline = new Pipeline().setStages(Array(document_assembler, word_segmenter , lemmatizer))
val data = Seq("이렇게되면이러한인간형을다투어본받으려할것이틀림없다.").toDF("text")
val result = pipeline.fit(data).transform(data)
import nlu

text = ["이렇게되면이러한인간형을다투어본받으려할것이틀림없다."]
lemma_df = nlu.load('ko.lemma').predict(text, output_level = "document")


{'lemma': [Annotation(token, 0, 2, 이렇게, {'sentence': '0'}),
   Annotation(token, 3, 4, 되+면, {'sentence': '0'}),
   Annotation(token, 5, 7, 이러한+ㄴ, {'sentence': '0'}),
   Annotation(token, 8, 11, 인간형+을, {'sentence': '0'}),
   Annotation(token, 12, 15, 다투어본, {'sentence': '0'}),
   Annotation(token, 16, 18, 받으할, {'sentence': '0'}),
   Annotation(token, 18, 18, 려, {'sentence': '0'}),
   Annotation(token, 20, 21, 것+이, {'sentence': '0'}),
   Annotation(token, 22, 25, 틀림없+다, {'sentence': '0'}),
   Annotation(token, 26, 26, ., {'sentence': '0'})]}

Model Information

Model Name: lemma
Compatibility: Spark NLP 2.7.0+
Edition: Official
Input Labels: [document]
Output Labels: [token]
Language: ko

Data Source

The model was trained on the universal dependencies from Korea Advanced Institute of Science and Technology (KAIST) dataset.


  • Building Universal Dependency Treebanks in Korean, Jayeol Chun, Na-Rae Han, Jena D. Hwang, and Jinho D. Choi. In Proceedings of the 11th International Conference on Language Resources and Evaluation, LREC’18, Miyazaki, Japan, 2018.