Government data governance is undergoing a new phase of transition from ``physical data aggregation'' to ``logical semantic unification''. Thus far, long-term ``autonomy'' of government information silos, leads to a wide spectrum of metadata curation issues, such as attributes with the same names but having different meanings, or attributes with different names but having the same meanings. Instead of either rebuilding/modifying legacy information systems or physically aggregating data from government information silos, logical semantic unification solves this problem by unifying the semantic expression of the metadata in government information silos and achieves the standardized metadata governance. This paper focuses on the logical semantic unification that semantically aligns the metadata in each government information silo with the existing standard metadata. Specifically, the names of the standard metadata are abstracted as semantic labels, and the column projections of silo relational data are semantically recognized to semantically align column names with the standard metadata and ultimately achieve the standardized governance of silo metadata. The existing semantic recognition techniques based on column projection fail to capture the column order-independent features of relational data and the correlation features among attributes and semantic labels. To address the above problem, we propose a two-phase model based on a prediction phase and a correction phase. In the prediction phase, a Co-occurrence-Attribute-Interaction (CAI) model is proposed to guarantee the column order-independent property by employing the parallelized self-attention mechanism; in the correction phase, a correction mechanism is introduced to optimize the prediction results of the CAI model by utilizing the co-occurrence of semantic labels. Experiments are conducted on a government benchmark dataset and several public English datasets, such as Magellan, and the results show that the two-phase model with a correction mechanism outperforms the current optimal model in macro-average and weighted average by up to 20.03% and 13.36%, respectively.