Deep learning based approach to unstructured record linkage

Purpose In the world of big data, data integration technology is crucial for maximising capability data-driven decision-making. Integrating from multiple sources drastically expands power information and allows us to address questions that are impossible answer using a single source. Record Linkage (RL) task identifying linking records describe same real object (e.g. person), it plays role in process. RL challenging, as uncommon different share unique identifier. Hence, must be matched based on comparison their corresponding values. Most existing techniques assume across structured represented by scheme (i.e. set attributes). Given increasing amount heterogeneous sources, those assumptions rather unrealistic. The purpose this paper propose novel model unstructured data. Design/methodology/approach previous work (Jurek-Loughrey, 2020), authors proposed approach application Siamese Multilayer Perceptron model. It was demonstrated method performed par with other approaches make constraining regarding This originally presented at iiWAS2020 [16] exploring new architectures Neural Network, which improves generalisation makes less sensitive parameter selection. Findings experimental results confirm Autoencoder-based architecture Network obtains better (Jurek et al. , 2020). Better have been achieved three out four sets. Furthermore, has second (hybrid) integrating Autoencoder model, more stable terms Originality/value To problem RL, presents deep learning improve Preceptron