XLM-RoBERTa is a multilingual BERT variant that encodes text from different languages into a shared semantic vector space. section 3.3.3
Within COMET, XLM-RoBERTa supports cross-lingual comparison by:
- Independently encoding the source, candidate, and reference into embedding vectors. section 3.3.3
- Placing German, English, and other language texts in the same space so their meanings can be compared. section 3.3.3
- Providing embeddings that COMET combines into interaction features for predicting translation quality. section 3.3.3
Translation-quality example
For the German source “Die Bank war voller Kunden,” XLM-RoBERTa encodes the source, the English candidate “The bank was full of customers,” and the reference “The financial institution was crowded with clients” into comparable representations. section 3.3.3
The architecture figure shows independent cross-lingual encoding, pooling into sentence embeddings, and downstream COMET scoring or ranking. section 3.3.3
