STATISTICAL LANGUAGE AND NEURAL NETWORK MODELS: CLASSIFYING HUMAN INSTRUCTIONS IN SITUATED ROBOT COMMAND

Abstract

To support natural language understanding in research, we aim to build upon previous research that was done by Traum et al, Gervits et al., and Marge et al. where previous experiments were done to create a classifier-based approach to automate human-robot dialog in an urban search and rescue scenario. Their experiments led to collecting a corpus of data called SCOUT (Situated Corpus of Understanding Transactions) along with building a classifier using the NPCEditor platform to build an effective question-answering conversational system. In previous experiments, four in total, a remote autonomous robot was used to collect the data used to build the corpus. This robot and the environment in which the experiments occurred were physical in experiments 1 and 2 but simulated in experiments 3 and 4. Human volunteers termed “Commanders”, would interact with the robot via two Wizards of Oz (WOZ), a Dialog Manager (DM), and a Robot Navigator (RN), performing the task of getting commands and responses between the commander and the robot. We aim to use the aforementioned SCOUT data to train two machine-learning models using the Bag-Of-Words method for featurization, with a random forest classifier and with a Neural Network using word embeddings. Where the “commander” statements are features and the final commands given to the robot navigator, are the labels. The results from both experiments will be compiled and analyzed with the ultimate goal of eliminating the use of Wizard-of-Oz methods that were used in the original four experiments.