Riassunto analitico
Throughout the past decade we have all assisted to the rise of Artificial Intelligence, from a novelty to its widespread adoption, in no small measure thanks to the advancement of hardware capabilities, availability of large datasets and software frameworks that simplified their use. If AI systems were to be employed in anything more than a research laboratory, their performance and cost had to rival those of humans. Scientists raced to build technologies that could achieve better and better metrics and, in doing so, traditional Machine Learning algorithms were shadowed by Deep Learning. But now, these systems have surpassed human capabilities in several tasks, and with comprehensive legislation on the horizon it is becoming clear that black-box neural networks cannot be the panacea. In this research I will tackle the problem of Entity Matching (also referred to as Entity Resolution) and explore strategies to merge the capabilities of a transformer-based language model and those of an inherently explainable, additive model on such tasks. My experiments show that these additive models can provide an explanation that reflects the final prediction, with a negligible loss in performance in comparison to the bare language models.
|
