Alttrans

Alttrans is a framework and data construct used in translation studies and natural language processing to capture multiple plausible translations of a single source sentence or text segment.

The term is a portmanteau of alternative and translation. It emerged in academic discussions in the 2010s to address the variability in translation choices across dialects, registers, and language pairs. Alttrans contrasts with traditional single-reference translation sets used for automatic MT evaluation, which may underrepresent variation in real-world translation practices.

Core features include: support for multi-reference translations; alignment metadata linking each variant to the source; metadata such as register, domain, or tone; scoring and evaluation methods capable of handling multiple targets; and data formats for storing alttrans entries (for example, a variant list with language code, text, and tags). Alttrans is used to train and evaluate machine translation systems, to study lexical and stylistic variation, and to create more robust evaluation metrics that reflect translation diversity.

Example. Source: The bank will approve the loan. Alttrans: The bank is expected to approve the loan. Alttrans: The bank plans to approve the loan. Alttrans: The bank will authorize the loan.

Limitations and commentary include: it is not standardized across the field, with coverage and conventions varying by project; quality depends on annotators and guidelines; alignment and aggregation of multiple variants can be complex; there are privacy considerations when dealing with sensitive texts. See also multi-reference translation, translation studies, corpus linguistics, and evaluation metrics for machine translation.