Interpreting natural language actions in a simulated world is the first step towards robots controlled by natural language commands. In this work we present several models for interpreting unrestricted natural language commands in a simple block world. We present and compare rule-based models and recurrent neural network models of various architectures. We also discuss strategies to deal with errors in natural language data and compare them. On the Language Grounding dataset, our models outperform the previous state-of-the-art results in both source and location prediction reaching source accuracy 98.8% and average distance 0.71 between the correct and predicted location.
Identifer | oai:union.ndltd.org:nusl.cz/oai:invenio.nusl.cz:365181 |
Date | January 2017 |
Creators | Pišl, Bedřich |
Contributors | Mareček, David, Variš, Dušan |
Source Sets | Czech ETDs |
Language | English |
Detected Language | English |
Type | info:eu-repo/semantics/masterThesis |
Rights | info:eu-repo/semantics/restrictedAccess |
Page generated in 0.0019 seconds