Fast and Accurate Language Identification - 43 Languages (CNN)

Description

Language detection and identification is the task of automatically detecting the language(s) present in a document based on the content of the document. LanguageDetectorDL is an annotator that detects the language of documents or sentences depending on the inputCols. In addition, LanguageDetetorDL can accurately detect language from documents with mixed languages by coalescing sentences and select the best candidate.

We have designed and developed Deep Learning models using CNNs in TensorFlow/Keras. The models are trained on large datasets such as Wikipedia and Tatoeba with high accuracy evaluated on the Europarl dataset. The output is a language code in Wiki Code style: https://en.wikipedia.org/wiki/List_of_Wikipedias.

This model can detect the following languages:

Arabic, Belarusian, Bulgarian, Czech, Danish, German, Greek, English, Esperanto, Spanish, Estonian, Persian, Finnish, French, Hebrew, Hindi, Hungarian, Interlingua, Indonesian, Icelandic, Italian, Japanese, Korean, Latin, Lithuanian, Latvian, Macedonian, Marathi, Dutch, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Serbian, Swedish, Tagalog, Turkish, Tatar, Ukrainian, Vietnamese, Chinese.

Predicted Entities

ar, be, bg, cs, da, de, el, en, eo, es, et, fa, fi, fr, he, hi, hu, ia, id, is, it, ja, ko, la, lt, lv, mk, mr, nl, pl, pt, ro, ru, sk, sl, sr, sv, tl, tr, tt, uk, vi, zh.

Open in Colab Download

How to use

...
language_detector = LanguageDetectorDL.pretrained("ld_wiki_tatoeba_cnn_43", "xx")\
   .setInputCols(["sentence"])\
   .setOutputCol("language")
languagePipeline = Pipeline(stages=[documentAssembler, sentenceDetector, language_detector])
light_pipeline = LightPipeline(languagePipeline.fit(spark.createDataFrame([['']]).toDF("text")))
result = light_pipeline.fullAnnotate("Spark NLP est une bibliothèque de traitement de texte open source pour le traitement avancé du langage naturel pour les langages de programmation Python, Java et Scala.")
...
val languageDetector = LanguageDetectorDL.pretrained("ld_wiki_tatoeba_cnn_43", "xx")
   .setInputCols("sentence")
   .setOutputCol("language")
val pipeline = new Pipeline().setStages(Array(documentAssembler, sentenceDetector, languageDetector))
val result = pipeline.fit(Seq.empty["Spark NLP est une bibliothèque de traitement de texte open source pour le traitement avancé du langage naturel pour les langages de programmation Python, Java et Scala."].toDS.toDF("text")).transform(data)
import nlu

text = ["Spark NLP est une bibliothèque de traitement de texte open source pour le traitement avancé du langage naturel pour les langages de programmation Python, Java et Scala."]
lang_df = nlu.load('xx.classify.wiki_43').predict(text, output_level='sentence')
lang_df

Results

'fr'

Model Information

Model Name: ld_wiki_tatoeba_cnn_43
Compatibility: Spark NLP 2.7.0+
Edition: Official
Input Labels: [sentence]
Output Labels: [language]
Language: xx

Data Source

Wikipedia and Tatoeba

Benchmarking

Evaluated on Europarl dataset which the model has never seen:

+--------+-----+-------+------------------+
|src_lang|count|correct|         precision|
+--------+-----+-------+------------------+
|      fr| 1000|   1000|               1.0|
|      nl| 1000|    999|             0.999|
|      sv| 1000|    999|             0.999|
|      pt| 1000|    999|             0.999|
|      it| 1000|    999|             0.999|
|      es| 1000|    999|             0.999|
|      fi| 1000|    999|             0.999|
|      el| 1000|    998|             0.998|
|      de| 1000|    997|             0.997|
|      da| 1000|    997|             0.997|
|      en| 1000|    995|             0.995|
|      lt| 1000|    986|             0.986|
|      hu|  880|    867|0.9852272727272727|
|      pl|  914|    899|0.9835886214442013|
|      ro|  784|    765|0.9757653061224489|
|      et|  928|    899|           0.96875|
|      cs| 1000|    967|             0.967|
|      sk| 1000|    966|             0.966|
|      bg| 1000|    960|              0.96|
|      sl|  914|    860|0.9409190371991247|
|      lv|  916|    856|0.9344978165938864|
+--------+-----+-------+------------------+

+-------+--------------------+
|summary|           precision|
+-------+--------------------+
|  count|                  21|
|   mean|  0.9832737168612825|
| stddev|0.020064155103808722|
|    min|  0.9344978165938864|
|    max|                 1.0|
+-------+--------------------+