Yes, Sentence Transformers can handle languages other than English. Multilingual sentence embeddings are achieved through models specifically trained to map text from multiple languages into a shared semantic space. These models are designed to understand and represent the meaning of sentences across languages, enabling tasks like cross-lingual similarity search or clustering without requiring translation. For example, a sentence in French and its English translation would produce embeddings that are close to each other in the vector space, even though the surface-level text differs.
The key to multilingual capability lies in training data and model architecture. Models like paraphrase-multilingual-MiniLM-L12-v2 or distiluse-base-multilingual-cased are trained on parallel corpora—collections of text translated into multiple languages (e.g., Wikipedia articles, multilingual news datasets). During training, the model learns to align embeddings of semantically equivalent sentences across languages. This is often done using contrastive learning objectives, such as triplet loss, where the model minimizes the distance between embeddings of translations while maximizing it for unrelated sentences. Additionally, some models leverage techniques like language-agnostic tokenization (e.g., using SentencePiece or multilingual BERT-style tokenizers) to handle diverse scripts and vocabularies. For instance, XLM-R (a common backbone for multilingual models) uses a single vocabulary trained on 100+ languages, enabling consistent tokenization across them.
From a practical standpoint, developers can use pre-trained multilingual Sentence Transformers off-the-shelf. For example, sentence-transformers/all-MiniLM-L6-v2 supports 50+ languages. When processing text, the model tokenizes input using its built-in tokenizer, which splits text into subwords common across languages. The embeddings generated will reflect semantic meaning regardless of language, enabling applications like multilingual search engines or multilingual document classification. However, performance may vary depending on the language and training data coverage—languages with abundant training data (e.g., German, Spanish) often yield better results than low-resource languages. Fine-tuning on domain-specific multilingual data can further improve accuracy for specialized use cases.