Table parsing is a key task in document understanding, consisting mainly of three core subtasks: table structure prediction, cell position prediction, and cell content prediction. However, Chinese engineering tables often contain hundreds of cells and require converting complex tabular structure information into a lengthy sequence representation during parsing, which makes it challenging for models to achieve ideal performance. In addition, these three subtasks are typically treated in isolation during modeling, and current research has not fully explored their tight spatial and semantic interdependencies. To address these challenges, this paper proposes a Local Attention (LA) and Semantic-Aware Unified Decoder-based (SAUD) framework for complex Chinese engineering table parsing—called LASDTab. To mitigate the difficulties caused by long tables, we incorporate a local attention mechanism into the structure decoder, improving the accuracy of long-table decoding tasks. Concerning the issue of separating multiple subtasks, we design SAUD’s shared attention mechanism and a semantic alignment mechanism to model cell positions and content jointly. In contrast, a dynamic triggering mechanism ensures consistency across subtasks. We compare the proposed LASDTab with six mainstream table parsing methods on four datasets. Experimental results demonstrate that LASDTab achieves significant improvements in TEDS, S-TEDS, and AP50, offering an efficient and robust solution to complex Chinese engineering table parsing tasks.