Failure-Aware Long-Form Translation: Design and Implementation of a Recoverable LLM Translation System
2026-08-10 • Computation and Language
Computation and Language
AI summaryⓘ
The authors studied problems with long translations that can look okay to an API but end up broken or incomplete. They created a step-by-step method to catch and fix these issues by checking early parts of the text and tracking if text gets replaced or continued. Their system only keeps interrupted translations if a part matches the original source and follows a set order and deadline for retrying, falling back when needed. They tested their method with different cases and showed it works as planned for controlling translation flow, though quality and detection still need separate testing.
long-form translationAPI layeroutput validationstream eventstranslation recoverytranslation fallbackprovenance markingcontrol flowinterrupted streamscompletion labels
Authors
Yanlin Yu
Abstract
A long-form translation request can succeed at the API layer and still produce an unusable result. The output may be empty, truncated, filtered, dominated by source or prompt material, or interrupted after producing text worth keeping. This report describes a recovery protocol developed for a deployed translation system with heterogeneous inputs and provider APIs. It delays the first visible release behind a 64-character window, validates the assembled output, and uses typed stream events to distinguish replacement from continuation. Interrupted work is retained only when a paragraph or sentence prefix can be re-derived from the source. Further attempts follow a stable model order and a shared deadline before entering a provenance-marked fallback path. A sanitized companion artifact implements the protocol and passes 38 public tests. Its fixed cases reproduce all 14 configured completion labels, contain four early-invalid prefixes before any of their 235 characters become visible, retain 31 boundary-safe characters across four interrupted streams, and satisfy the attempt, event, and provenance rules in two end-to-end scenarios. These results are executable checks of the published control flow. Translation quality and detector performance on naturally occurring outputs require a different evaluation.