Intent Detection and Semantic Parsing for Navigation Dialogue Language Processing

Voice-based human-machine interface has become a prevalent feature for modern intelligent vehicles, especially in navigation and infotainment applications. Automatic Speech Recognition (ASR) converts spoken audio streams to plain texts, but a follow-up Natural Language Processing (NLP) sub-system is needed to understand the contextual meaning from the text and act. For the specific human-vehicle navigation dialogue application, the two major tasks include (1) intent detection - decide whether a sentence is navigation-related, and (2) semantic parsing - retrieve important information (e.g., point-of-interest destinations) from the words. To address these two tasks, this study proposes a Recurrent Neural Network (RNN) architecture, with the consideration of (1) one joint model vs. two separate models, (2) a context window approach vs. sequence-to-sequence translation approach, as well as (3) alternate model hyper-parameter selections. The experiment is conducted with both the benchmark ATIS dataset and the CU-Move in-vehicle dialogue corpus, and the result is compared against related state-of-the-art methods. Overall, the proposed solution on the CU-Move data achieves accuracies of 98.24% for intent detection and 99.60% for semantic parsing, outperforming other related methods.