Understanding natural language sentences with word embedding and multi-modal interaction

Understanding and grounding human commands with natural languages have been a fundamental requirement for service robotic applications. Although there have been several attempts toward this goal, the bottleneck still exists to store and process the corpora of natural language in an interaction system. Currently, the neural- and statistical-based (N&S) natural language processing have shown potential to solve this problem. With the availability of large data-sets nowadays, these processing methods are able to extract semantic relationships while parsing a corpus of natural language (NL) text without much human design, compared with the rule-based language processing methods. In this paper, we show that how two N&S based word embedding methods, called Word2vec and GloVe, can be used in natural language understanding as pre-training tools in a multi-modal environment. Together with two different multiple time-scale recurrent neural models, they form hybrid neural language understanding models for a robot manipulation experiment.