Structured learning for spoken language understanding in human-robot interaction

Robots are slowly becoming apart of everyday life, being marketed for commercial applications such as telepresence, cleaning or entertainment. Thus, the ability to interact via natural language with non-expert users is becoming a key requirement. Even if user utterances can be efficiently recognized and transcribed by automatic speech recognition systems, several issues arise in translating them into suitable robotic actions and most of the existing solutions are strictly related to a specific scenario. In this paper, we present an approach to the design of natural language interfaces for human robot interaction, to translate spoken commands into computational structures that enable the robot to execute the intended request. The proposed solution is achieved by combining a general theory of language semantics, i.e. frame semantics, with state-of-the-art methods for robust spoken language understanding, based on structured learning algorithms. The adopted data driven paradigm allows the development of a fully functional natural language processing chain, that can be initialized by re-using available linguistic tools and resources. In addition, it can be also specialized by providing small sets of examples representative of a target newer domain. A systematic benchmarking resource, in terms of a rich and multi-layered spoken corpus has also been created and it has been used to evaluate the natural language processing chain. Our results show that our processing chain, trained with generic resources, provides a solid baseline for command understanding in a service robot domain. Moreover, when domain-dependent resources are provided to the system, the accuracy of the achieved interpretation always improves.