Comparative Analysis of Question Answering Models for HRI Tasks with NAO in Spanish

Recent studies on Human Robot Interaction (HRI) have shown that different types of applications that combine metrics and techniques can help achieve a more efficient and organic interaction. This applications can be related to human care or go further and use a humanoid robot for nonverbal communication tasks. Similarly, for verbal communication, we found Question Answering, a Natural Language Processing task, that is in charge of capturing and interpret a question automatically and return a good representation of an answer. Also, recent work on creating Question Answering models, based on the Transformer architecture, have obtained state-of-the-art results. Our main goal in this project is to build a new Human Robot Interaction technique which uses a Question Answering system where we will test with college students. In the creation of the Question Answering model, we get results from state-of-the-art pre-trained models like BERT or XLNet, but also multilingual ones like m-BERT or XLM. We train them with a new Spanish dataset translated from the original SQuAD getting our best results with XLM-R, obtaining 68.1 F1 and 45.3 EM in the MLQA test dataset, and, 77.9 F1 and 58.3 EM for XQuAD test dataset. To validate the results obtained, we evaluated the project based on HRI metrics and a survey. The results demonstrate a high degree of acceptance in the students about the type of interaction that has been proposed.