Hybrid Artificial Intelligence to extract patterns and rules from argumentative and legal texts

This Thesis is composed of a selection of studies realized between 2019 and 2022, whose aim is to find working methodologies of Artificial Intelligence (AI) and Machine Learning for the detection and classification of patterns and rules in argumentative and legal texts. We define our approach as
“hybrid”, since different methods have been employed combining symbolic AI (which involves “top-dow” structured knowledge) and sub-symbolic AI (which involves “bottom-up” data-driven knowledge).
The first group of these works was dedicated to the classification of argumentative patterns. Following the Waltonian model of argument (according to which arguments are composed by a set of premises and a conclusion), and the theory of Argumentation Schemes, this group of studies was focused on the detection of argumentative evidences of support and opposition. More precisely, the aim of these first works was to show that argumentative patterns of opposition and support could be classified at fine-grained levels and without resorting to highly engineered features. To show this, we firstly employed methodologies based on Tree Kernel classifiers and TFIDF. In these experiments, we explored different combinations of Tree Kernel calculation and different data structures (i.e., different tree structures). Also, some of these combinations employs a hybrid approach where the calculation of similarity among trees is influenced not only by the tree structures but also by a semantic layer (e.g. those using “smoothed” trees and “compositional” trees).
After the encouraging results of this first phase, we explored the use of a new methodology which was deeply changing the NLP landscape exactly in that year, fostered and promoted by actors like Google, i.e. Transfer Learning and the use of language models. These newcomer methodologies markedly improved our previous results and provided us with stronger NLP tools. Using Transfer Learning, we were also able to perform a Sequence Labelling task for the recognition of the exact span of argumentative components (i.e. claims and premises), which is crucial to connect the sphere of natural language to the sphere of logic.
The last part of this work was finally dedicated to show how to use Transfer Learning for the detection of rules and deontic modalities. In this case, we tried to explore a hybrid approach which combines structured knowledge coming from two LegalXML formats (i.e., Akoma Ntoso and LegalRuleML) with sub-symbolic knowledge coming from pre-trained (and then fine-tuned) neural architectures.