Despite major advances in machine translation (MT) in recent years, progress remains limited for many low-resource languages that lack large-scale training data and linguistic resources. In this paper, we introduce SINITICMTERROR, a novel fine-grained dataset that builds on existing parallel corpora to provide error span, error type, and error severity annotations in machine-translated examples from English to Mandarin, Cantonese, and Wu Chinese, along with a Mandarin-Hokkien component derived from a non-parallel source. Our dataset serves as a resource for the MT community to fine-tune models with error detection capabilities, supporting research on translation quality estimation, error-aware generation, and low-resource language evaluation. We also establish baseline results using language models to benchmark translation error detection performance. Specifically, we evaluate multiple open source and closed source LLMs using span-level and correlation-based MQM metrics, revealing their limited precision, underscoring the need for our dataset. Finally, we report our rigorous annotation process by native speakers, with analyses on pilot studies, iterative feedback, insights, and patterns in error type and severity.

Despite the growing adoption of large language models (LLMs) for machine translation (MT) evaluation, reliably detecting fine-grained error spans in low-resource languages remains challenging. A key limitation is that LLMs often lack sufficient cross-lingual knowledge, making it difficult for them to accurately localize errors at the span level. In this work, we introduce Cross-lingual Semantic Scaffolding, a two-part prompting framework that incorporates sentence-level parallel examples and word-level lexical mappings at inference time. We evaluate this approach on five LLMs using controlled ablation experiments on Taiwanese Hokkien. The proposed framework consistently improves performance across all tested models, increasing average Span F1 by +0.175 (+33.6%) and average Span Recall by +0.351 (+75.9%), and shows that the gains are mainly driven by improved error coverage. It also supports AI-assisted human annotation, reducing expert annotation time from 450 seconds to 90.92 seconds per sentence, with 98.2% of AI-proposed spans accepted without modification. These results demonstrate that Cross-lingual Semantic Scaffolding is both an effective detection method and a practical tool for reducing expert annotation costs in low-resource MT evaluation.