Improved Spoken Language Representation for Intent Understanding in a Task-Oriented Dialogue System.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Speech Recognition

Background:

• Deep learning has advanced text-based intent classification.
• Automatic speech recognition (ASR) errors in spoken dialogue systems can lead to misclassified user intents.
• Existing methods struggle with the inaccuracies introduced by ASR.

Purpose of the Study:

• To propose a novel approach for robust intent classification in spoken dialogue systems.
• To address the challenge of ASR errors impacting intent recognition.
• To enhance the performance of intent classification despite noisy speech input.

Main Methods:

• A novel approach jointly utilizing ASR-recognized text and labeled text.
• Development of a fine-tuned recognized language model (RLM) for evaluation.
• Integration of speech recognition outputs with language understanding models.

Main Results:

• The proposed scheme demonstrates effectiveness in classifying intents within spoken dialogue systems.
• The method successfully mitigates the negative impact of ASR errors on intent classification accuracy.
• Experimental results validate the proposed approach's performance.

Conclusions:

• The novel approach significantly improves intent classification accuracy in spoken dialogue systems with ASR errors.
• Jointly using recognized text and labeled data offers a robust solution.
• The fine-tuned RLM is effective for intent classification in real-world spoken interactions.