Comprendere e rispondere a comandi in linguaggio naturale: Human Computation e Reti Neurali ​

The aim of the work on which this thesis is based is to develop a system able to understand commands and intents from their natural language form. The first approach used to reach the goal was based on Crowdsourcing and Human Computation. The reason behind it was the idea that people have an innate ability to understand the meaning of a phrase regardless from the actual words used. This approach has been tested as showed in the thesis and its limits have been discussed. To improve upon the first approach a second one, using pre-trained neural networks as the Universal Sentence Encoder, has been adopted. The first goal of this new approach was to simplify the tasks given to the workers enough to be solvable by non-experts, but some tests regarding the standalone use of this method have also been carried out and discussed. To conclude both approaches have been tested and analysed exposing their limits and advantages.​