A carregar...
Publicação
Artificial Intelligence for The Automatic Identification of Addresses: The Portuguese Case
Orientador(es)
Resumo(s)
Addresses can be used as quasi-identifiers to link relevant data across multiple registers, in a process known as address matching, essential to activities like urban planning, location-based services, and administrative census operations. On the other hand, address data quality has a direct impact on demographic and other spatial analyses, since it may lead to uncertainty and potential bias. This thesis aims at contributing to current knowledge in this field, using residential addresses managed by Statistics Portugal as a case study. We start by proposing a multiclass classification algorithm to evaluate the syntactic quality of residential addresses from a large database managed by Statistics Portugal, based on the NSGA-II algorithm and two modified kNN algorithms. The results show improved classification performance over baseline methods while simultaneously delivering insights on relevant features and local patterns, without resorting to external databases, one of the limitations found in similar studies. Regarding record linkage, we adopt deep learning models for semantic address matching, namely pretrained language models such as BERT or one of its variants. We train a BERT-based model from scratch and compare it with a novel vocabulary-free approach, based on ByT5, which shows competitive results in terms of accuracy. To further optimize models’ implementation, we adopt strategies such as automatic labelling of training datasets combined with synthetic datasets generation and in-batch negatives loss to minimize human effort and optimization methods such as automatic mixed precision to reduce computational overhead. As further developments, we propose the use of new generation transformer-based models, a privacy-preserving temporal record linkage approach and the optimization of algorithms using evolutionary coding agents such as AlphaEvolve.
Descrição
A thesis submitted in partial fulfillment of the requirements for the degree of Doctor in Information Management
Palavras-chave
Automated Address Matching Address Parsing Administrative Census Deep Learning NLP