Extracting Frame-based Knowledge Representation from Route Instructions

To build a simulated robot that follows route instructions in unconstrained natural language, we propose a frame-segment decoding algorithm that achieves two joint tasks: (1) Segment the route instruction so that the sequence of segments corresponds to a sequence of actions required to complete the task, and (2) Choose the slot fillers of the frame-based knowledge representation for each action. Our model was trained with voted perceptron. We also created labels for actions by merging the slot fillers of their frame-based knowledge representation, reducing the problem to a sequence labeling task. As a comparison, we applied a linear-chain Conditional Random Fields to this representation. The experimental result shows that our model performs better with a 77.7% success rate.