Automatic Generation of German Translation Candidates for SNOMED CT Textual Descriptions

Area of Science:

• Medical Informatics
• Computational Linguistics
• Natural Language Processing

Background:

• SNOMED CT (Systematized Nomenclature of Medicine - Clinical Terms) is a comprehensive clinical terminology.
• Localization of SNOMED CT into different languages is crucial for global adoption and use.
• Existing machine translation tools may not provide sufficient accuracy for specialized terminologies like SNOMED CT.

Purpose of the Study:

• To introduce and evaluate the Multiple Translation Paths (MTP) approach for assisting human translation in SNOMED CT localization.
• To assess the effectiveness of combining multiple free, web-based machine translation tools and language combinations.
• To determine if a majority voting system on multiple translation candidates improves translation quality.

Main Methods:

• Developed the MTP approach, which generates scored translation candidates (TCs) for each input SNOMED CT concept.
• Utilized free machine translation engines (Google Translator, DeepL, Systran) with multiple source languages (English, Spanish, Swedish, French) and support languages.
• Applied MTP to the SNOMED CT Starter Set, generating up to 91 translation results per concept.
• Conducted a descriptive assessment of TC variety and analyzed typical results, focusing on German as the target language.

Main Results:

• The MTP approach capitalizes on combinatorial growth by combining input languages, support languages, and translation engines.
• Analysis revealed that many translation candidates are derived via different translation paths, indicating redundancy.
• Combinations of translation engines produced distributions with a higher number of distinct TCs per concept.
• Initial qualitative analysis showed promising results, supporting the hypothesis that majority voting enhances translation quality.

Conclusions:

• The MTP approach offers a viable strategy for enhancing machine translation in SNOMED CT localization projects.
• Leveraging multiple free MT tools and diverse language combinations can yield a richer set of translation candidates.
• Majority voting among multiple translation candidates shows potential for improving translation accuracy compared to single-engine outputs.