NLP / AI Engineer for DeepL Glossary Creation & Translation Quality Testing

We are looking for a freelancer to help us build and validate multilingual translation glossaries using a large dataset of existing product translations. Project background We maintain product information in a PIM system. English is our source language, and our product content has historically been translated into multiple languages. We can provide exports from our PIM containing: - English source text - Existing translated versions - Product names and descriptions - Product attributes and other translatable properties - Large numbers of existing product records that can be used as training/reference data Our goal is to analyse these existing translations and use them to create high-quality glossaries that can eventually be used with DeepL as part of an automated translation workflow. What we need 1. Analyse our existing translations Use AI, NLP, scripts, or other suitable tools to analyse the supplied multilingual dataset and identify: - Company and brand-specific terminology - Product-specific and industry-specific terminology - Terms that should consistently be translated in a particular way - Terms, brand names, product ranges, materials, colours, etc. that should not be translated - Recurring translation patterns - Existing terminology conventions - Language-specific conventions - Cases where translations depend on context - Inconsistencies or potentially incorrect translations in the historical data We are not looking to simply extract frequently occurring words. The objective is to understand how terminology is actually being used across our product catalogue and identify rules that are suitable for a translation glossary. 2. Create translation glossaries Using the findings above, create structured glossaries suitable for use with DeepL. The glossaries should ideally include: - Source term - Target term - Source/target language - Examples or context where useful - Confidence/frequency information where appropriate - Notes explaining special cases or terminology rules We have multiple target languages, so the approach needs to be repeatable rather than manually building each glossary from scratch. 3. Build a glossary validation/test process An important part of the project is proving that the glossaries actually improve translation quality. We would like a repeatable test system that can: - Take existing English product information. - Translate it using DeepL without the glossary. - Translate it using the newly created glossary. - Compare both results against our existing human translations. - Highlight differences and recurring patterns. - Identify where the glossary improves or worsens the translation. - Produce useful metrics/reports to help us refine the glossary. We are particularly interested in techniques that go beyond simple exact string comparison, as a translation can differ from our existing translation while still being perfectly valid. The test process should help identify systematic divergence, terminology problems and glossary rules that may be producing unexpected results. 4. Human-assisted validation Once the automated validation is producing good results, we would like to perform human review on a representative/random sample of products. The results should allow us to iteratively: analyse → build glossary → test → identify problems → refine glossary → human validate Longer-term objective: The final system will eventually be integrated into an existing Node.js application. The application will trigger translation of English product data from our PIM into multiple languages, using the approved glossaries to maintain consistent terminology. You do not necessarily need to build the production Node.js integration as part of this project, but an understanding of APIs and application integration would be useful. Skills we are looking for Important: - Experience with machine translation / translation technology - Experience analysing multilingual datasets - AI / LLM experience - Data analysis and scripting, ideally Python or similar - Experience working with large datasets - Understanding of terminology extraction and glossary creation - Ability to design a reliable translation evaluation/testing methodology Highly desirable: - DeepL API and DeepL Glossary experience - Experience with PIM/product catalogue data - NLP or computational linguistics experience - Experience evaluating machine translation quality - Experience using embeddings, LLMs or semantic similarity for translation analysis Bonus: - Node.js / TypeScript - API integration experience - E-commerce or product information management experience Deliverables We expect the project to produce: - Analysis of our historical translation data - Proposed terminology/glossary extraction methodology - Glossaries for the agreed target languages - A repeatable process/script for generating or updating glossaries - Automated translation test/validation platform - Comparison of translations with and without glossaries - Reporting of terminology mismatches and recurring translation problems - Human-validation workflow/sample - Documentation explaining how the system can later be integrated into our Node.js translation application When applying Please include: - Examples of similar translation, NLP or terminology projects you have worked on - Your experience with DeepL, if applicable - How you would approach extracting glossary terms from an existing multilingual product dataset - How you would evaluate whether a glossary is actually improving translation quality - Technologies/tools you would use - Any experience working with PIM or product catalogue data Please do not submit a generic translation proposal. This is primarily a technical data-analysis and machine-translation project rather than a manual translation project.