Multilingual Idiom Semantic Search and Translation System
Overview
This service provides a semantic search system based on vector representations (embeddings) designed to support multilingual idiom understanding and translation. The system converts English idioms into numerical vector representations that capture their semantic meaning. These representations enable the system to identify idioms with similar meanings across languages. When a user enters an idiom, the system computes its vector embedding and searches for the most semantically similar idioms using cosine similarity. The system then checks whether a direct equivalent idiom exists in Slovene. If a direct idiomatic translation exists, it is returned to the user. If not, the system provides the most appropriate semantic explanation of the idiom. The system is designed to support bidirectional functionality, allowing searches from English to Slovene as well as from Slovene to English. The service uses the SentenceTransformers framework and the multilingual-e5-large model to generate multilingual embeddings that allow comparison of semantic similarity across languages. Vector representations are compared using cosine similarity to enable cross-lingual semantic matching between English and Slovene idiomatic expressions.
Outcome
Users can obtain direct idiomatic equivalents between English and Slovene where they exist, or receive meaningful explanations when no direct equivalent is available. The service improves access to idiomatic language knowledge and supports language learning, translation, and linguistic research by enabling semantic search across languages.
Service Card
| Category | AI products |
| Domain | Multilingual NLP |
| Vertical | Digital Society |
| Target users | Translators and Language Professionals, Language Learners and Teachers, Linguists and NLP Researchers, Developers of Language Technologies, Students in Linguistics and Translation Studies |
| Service type | Model |
| Outputs | RAG vector database with a selected off-the-shelf LLM model for training; chat system for inference. |
| Requirements | Requires NVIDIA GPU node. |
| Limitations | The quality of results depends on the coverage of idioms in the dataset and the performance of multilingual embedding models. |
| Availability | Q1/2027 |
| Licensing | MIT and CC BY-NC-SA 4.0 |
| Visual content | |
| Visual caption |