Universal Sentence Encoder Multilingual Large


The Universal Sentence Encoder encodes text into high-dimensional vectors that can be used for text classification, semantic similarity, clustering, and other natural language tasks.

The model is trained and optimized for greater-than-word length text, such as sentences, phrases, or short paragraphs. It is trained on a variety of data sources and a variety of tasks with the aim of dynamically accommodating a wide variety of natural language understanding tasks. The input is the variable-length text and the output is a 512-dimensional vector. The universal-sentence-encoder model has trained with a deep averaging network (DAN) encoder.

This model supports 16 languages (Arabic, Chinese-simplified, Chinese-traditional, English, French, German, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Spanish, Thai, Turkish, Russian) text encoder.

The details are described in the paper “Multilingual Universal Sentence Encoder for Semantic Retrieval”.

Note: This model only works on Linux and macOS operating systems and is not compatible with Windows due to the incompatibility of the SentencePiece library.


How to use

embeddings = UniversalSentenceEncoder.pretrained("tfhub_use_multi_lg", "xx") \
      .setInputCols("document") \
val embeddings = UniversalSentenceEncoder.pretrained("tfhub_use_multi_lg", "xx")


It gives a 512-dimensional vector of the sentences.

Model Information

Model Name: tfhub_use_multi_lg
Compatibility: Spark NLP 2.7.0+
License: Open Source
Edition: Official
Input Labels: [sentence]
Output Labels: [sentence_embeddings]
Language: xx

Data Source

This embeddings model is imported from https://tfhub.dev/google/universal-sentence-encoder-multilingual-large/3


  • We apply this model to the STS benchmark for semantic similarity. The eval can be seen in the example notebook made available. Results are shown below:
STSBenchmark                       | dev    | test  |
Correlation coefficient of Pearson | 0.837  | 0.825 |
Dataset                | Quora | AskUbuntu | Average |
Mean Averge Precision  | 89.1  | 42.3      | 65.7    |
  • For the translation pair retrieval, we evaluate the model on the United Nation Parallal Corpus. Results are shown below:
Language Pair  | en-es  | en-fr | en-ru | en-zh |
Precision@1    | 86.1   | 83.3  | 88.9  | 78.8  |