• common knowledge
• common sense
• diagnostic dataset
• natural language inference
• natural language processing
• nli
• nlp
• probing task

Nowadays it has become more and more important to understand how much the neural models applied on Natural Language Processing can understand about language features. The standard method to address this kind of problem is to create some probing task in order to investigate the knowledge learned by pre-trained models for the Natural Language Inference task. The purpose of this thesis is to determine to which extent the models are able to address linguistic features such as the common sense and the common knowledge.​
We created a new data-set with 1000 couples of sentences regarding these kind of phenomena, defining a set of fine-grained categories to better describe the linguistic features. We tagged every pair of sentences with the expected label and we used the data-set to test different neural networks in order to analyze which kind of sub-phenomena they are able to understand.