Interactive Learning of Grounded Verb Semantics towards Human-Robot Communication

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1634–1644 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1150 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1634–1644 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1150 Interactive Learning of Grounded Verb Semantics towards Human-Robot Communication Lanbo She and Joyce Y. Chai Department of Computer Science and Engineering Michigan State University East Lansing, Michigan 48824, USA {shelanbo, jchai}@cse.msu.edu Abstract To enable human-robot communication and collaboration, previous works repre- sent grounded verb semantics as the poten- tial change of state to the physical world caused by these verbs. Grounded verb semantics are acquired mainly based on the parallel data of the use of a verb phrase and its corresponding sequences of primitive actions demonstrated by hu- mans. The rich interaction between teach- ers and students that is considered impor- tant in learning new skills has not yet been explored. To address this limitation, this paper presents a new interactive learning approach that allows robots to proactively engage in interaction with human partners by asking good questions to learn mod- els for grounded verb semantics. The pro- posed approach uses reinforcement learn- ing to allow the robot to acquire an op- timal policy for its question-asking be- haviors by maximizing the long-term re- ward. Our empirical results have shown that the interactive learning approach leads to more reliable models for grounded verb semantics, especially in the noisy environ- ment which is full of uncertainties. Com- pared to previous work, the models ac- quired from interactive learning result in a 48% to 145% performance gain when ap- plied in new situations. 1 Introduction In communication with cognitive robots, one of the challenges is that robots do not have sufficient linguistic or world knowledge as humans do. For example, if a human asks a robot to boil the wa- ter but the robot has no knowledge what this verb phrase means and how this verb phrase relates to its own actuator, the robot will not be able to exe- cute this command. Thus it is important for robots to continuously learn the meanings of new verbs and how the verbs are grounded to its underlying action representations from its human partners. To support learning of grounded verb seman- tics, previous works (She et al., 2014; Misra et al., 2015; She and Chai, 2016) rely on multiple in- stances of human demonstrations of correspond- ing actions. From these demonstrations, robots capture the state change of the environment caused by the actions and represent verb semantics as the desired goal state. One advantage of such state-based representation is that, when robots en- counter the same verbs/commands in a new situa- tion, the desired goal state will trigger the action planner to automatically plan a sequence of prim- itive actions to execute the command. While the state-based verb semantics provides an important link to connect verbs to the …

Content truncated to 500 words