Word Embeddings for Urdu (urduvec_140M_300d)

Description

This model is trained using Word2Vec approach on a corpora of 140 Million tokens, has a vocabulary of 100k unique tokens, and gives 300 dimensional vector outputs per token. The output vectors map words into a meaningful space where the distance between the vectors is related to semantic similarity of words.

These embeddings can be used in multiple tasks like semantic word similarity, named entity recognition, sentiment analysis, and classification.

Live Demo Open in Colab Download Copy S3 URI

How to use

...
embeddings = WordEmbeddingsModel.pretrained("urduvec_140M_300d", "ur") \
.setInputCols(["document", "token"]) \
.setOutputCol("embeddings")
nlp_pipeline = Pipeline(stages=[document_assembler, sentence_detector, tokenizer, embeddings])
pipeline_model = nlp_pipeline.fit(spark.createDataFrame([[""]]).toDF("text"))
result = pipeline_model.transform(spark.createDataFrame([['مجھے سپارک این ایل پی پسند ہے۔']], ["text"]))

val embeddings = WordEmbeddingsModel.pretrained("urduvec_140M_300d", "ur")
.setInputCols(Array("document", "token"))
.setOutputCol("embeddings")
val pipeline = new Pipeline().setStages(Array(document_assembler, sentence_detector, tokenizer, embeddings))
val data = Seq("مجھے سپارک این ایل پی پسند ہے۔").toDF("text")
val result = pipeline.fit(data).transform(data)

import nlu

text = ["مجھے سپارک این ایل پی پسند ہے۔"]
urduvec_df = nlu.load('ur.embed.urdu_vec_140M_300d').predict(text, output_level="token")
urduvec_df

Results

The model gives 300 dimensional Word2Vec feature vector outputs per token.

|Embeddings vector                                   |  Tokens 
|----------------------------------------------------|---------
| [0.15994004905223846, -0.2213257998228073, 0.0...  |	مجھے
| [-0.16085924208164215, -0.12259697169065475, -...  |	سپارک
| [-0.07977486401796341, -0.528775691986084, 0.3...  |	این
| [-0.24136857688426971, -0.15272589027881622, 0...  |	ایل
| [-0.23666366934776306, -0.16016320884227753, 0...  |	پی
| [0.07911433279514313, 0.05598200485110283, 0.0...  |	پسند

Model Information

Data Source

The model is imported from http://www.lrec-conf.org/proceedings/lrec2018/pdf/148.pdf