Semantic-aware data imputation in dynamic relational databases via pre-trained language models

• We formally define the Dynamic Data Imputation (DDI) problem on dynamic databases. • We introduce a semantic-aware metric capturing contextual similarity during imputation. • We propose SENTI, the first incremental algorithm for dynamic imputation based on Pretrained Language Models (PLMs). • SENTI is model-agnostic, efficient, and adaptable across datasets and PLM implementations. • Experiments show that SENTI is up to 100 times faster and 12% more accurate than baselines. Digital systems for information and representation management rely on database architectures, whose effectiveness is undermined by the presence of missing values. Data Imputation (DI) is a well-known process that replaces missing values, usually represented by means of nulls, with reliable constants. However, existing methods typically assume a static view of the database, overlooking the fact that real-world databases are often updated over time through the addition of new (possibly incomplete) information. We address Dynamic Data Imputation (DDI), that is the problem of imputing nulls in incrementally updated databases. We show that existing learning-based approaches are ill-suited for DDI, as they require costly retraining whenever the data increases over time. Instead, we propose a novel incremental algorithm called SENtence Transformer based Imputation ( SENTI ) that uses advanced techniques to perform quick and accurate similarity search by exploiting the inference capabilities of Pretrained Language Models, without any need for training. The experiments reveal that our technique outperforms the state-of-the-art static DI approaches (adapted to solve DDI) both in effectiveness and efficiency.