Natural Language Understanding For Human-Robot Interaction

It is inevitable to meet the human-robot interaction when a robot is at the service of a human. In an appropriate human-robot interaction, not only does the robot under-stand the human better but the human also realizes the status of the robot. However, to reach the natural human-robot interaction, it is necessary for the robot to solve the challenges such as the action capability, the versatility, and the comprehensive ability. In this dissertation, the author propose to utilize the human knowledge to enhance the robots’abilities of accomplishing the unforeseen tasks. It helps to solve the versatility of a robot. The precondition for a robot to utilize human knowledge is being able to understand the knowledge. After all, it requires the robot understanding natural lan-guage. At the mean while, speaking natural language is the most natural way for a human to interact with a robot. Hence, understanding natural language is the essential requirement to human-robot interaction. In order to sovle the challenges arose from natural human-robot interaction, this dissertation focuses on the large-scale natural lan-guage understanding. Three issues are step-by-step addressed and solved to improve the robots’abilities of understanding natural language.The first issue is how to extract the exact semantics of the open knowledge mainly in the form of natural language. While the robot designer could not predict all possible situations, robots need to dynamically acquire open knowledge to fill the knowledge gap to accomplish a task. To understand the knowledge in the form of natural language, a robot needs to understand the natural language. Meanwhile, these knowledge may be in multiple modes, e.g., semi-structured or unstructured (in natural language). Therefore, the multi-mode natural language processing techique is proposed. It employ an unified and general mechanism of semantic parsing to extract the deep semantics from the open knowledge in multiple modes. Furthermore, in such knowledge, some of the semantic informations are usually hidden due to presuppositions. Hence, the semantic retrieval for presupposition is also introduced to make sure the results of semantic extraction are complete and correct to capture the meaning of the original knowledge.The second is the unified formal representation for action knowledge and conver-sational utterance in human-robot interaction and the problem of understanding natural language on a large scale. In human-robot interaction, a robot needs to be able to under-stand the natural language in aciton knowledge and conversational utterance. General semantic parsing requires an unified formal representation to capture the meanings of these two different types of natural language. For this purpose, the human-robot dialog representation is proposed. And in the case of large-scale natural language understand-ing, the semantic parsing has to deal with the uncertainties and ambiguities. Hence, the probabilistic semantic parsing is employed. Another advantage of probabilistic seman-tic parsing is that, through a supervised learning algorithm, a semantic lexicon could be learned from the training examples annotated with semantic forms. It avoids the manual construction of a semantic lexicon.The final issue is how to acquire a hurge number of training examples with an-notations. Understanding natural language on a large scale needs the lexicon of a se-mantic parser covering enough words in a specific domain. However, learning a wide-coverage semantic parser usually requires a hurge number of training examples anno-tated with semantic forms. It is intractable to get such corpus due to the exhaustive time-consuming. To solve this problem, the lexicon propagation is proposed. It is an approach to wide-coverage semantic parsing by extending a semantic lexicon from low-coverage to wide-coverage through a large number of unannotated sentences. Lexicon propagation only requires a small number of annotated training examples to learn an ini-tial lexicon. Through lexicon propagation, the unannotated sentences help a semantic parser to reach wide-coverage. The proposed approach solves the problem of exhaustive annotations.