A Data Annotation Method for Fusion of Semantic and Physical Features

In large enterprises like the State Grid, data is often stored in two-dimensional tables, with column headers summarizing each column's meaning. However, operational errors or poor data management can result in missing headers, complicating data interpretation and usage. This issue is particularly critical for numeric columns, as their semantic type is essential for accurate interpretation. For instance, a numeric column without a header could represent equipment IDs, measurements, or statistical indicators, making it difficult to discern its meaning and affecting data analysis accuracy. To address this, we propose the MixTab model, a data annotation method that combines semantic features from non-numeric columns with physical features from numeric ones to detect semantic types. Tested on two popular datasets, Wikipedia and DBpedia, MixTab outperforms state-of-the-art methods. On the Wikipedia dataset, it improves numeric column header prediction by 3.2% in accuracy, 2.2% in recall, and 2.5% in F1 score. On the DBpedia dataset, it achieves gains of 2.7^%, 2.0%, and 1.9% in accuracy, recall, and F1 score, respectively.