A Neurosymbolic Framework for Markov Logic Networks

The impact of Deep Learning is due to the ability of its algorithm to mimic purely instinctive decisions, by reaching human-like performance in many isolated tasks, like image recognition or speech recognition. However, on the way to general artificial intelligence, which comprehend the set of cognitive abilities that gives machines a human-like intelligence it is extremely difficult to believe that these techniques, in an isolated way, can lead to a turning point. This because humans are still capable of performing more abstract and conscious reasoning processes on top of these instinctive tasks. This condition makes the need of a more complex and general theories, where the deep learning techniques constitutes only an ingredient of the final recipe. In this thesis, we propose an implementation of a neuro-symbolic framework to merge symbolic and sub-symbolic reasoning and we aim to investigate how this integration improves deep learning systems with the use of additional background knowledge in form of symbolic rules. Starting from Markov Logic Networks we introduce neural networks to provide different weights for different grounding of the same formula and we inject sub-symbolic capabilities into MLNs. Then, to test our implementation, we move to Argument Mining, a complex NLP task whose goal is the extraction of structured information from raw textual sources. We compare our approach to a baseline using only data without rules and to another neuro-symbolic framework, and establish that using logical rules during the training process gives a positive contribution to the task.