Joint intent detection and slot filling with syntactic and semantic features using multichannel CNN-BiLSTM

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Spoken language understanding is key for conversational agents, with intent detection and slot filling being core Natural Language Understanding (NLU) tasks.
• Joint learning of these tasks shows promise, but limited data hinders generalization.
• Traditional models struggle with semantic and syntactic nuances crucial for NLU.

Purpose of the Study:

• To propose a hybridized text representation for enhanced joint intent detection and slot filling.
• To address generalization challenges in NLU models using limited annotated datasets.
• To improve the accuracy and efficiency of conversational agents.

Main Methods:

• A multichannel convolutional neural network (CNN) integrating non-contextual (word2vec), part-of-speech (POS) tag, and contextual (BERT) embeddings.
• Shared bidirectional long short-term memory (BiLSTM) network processes combined embeddings.
• Separate softmax classifiers for intent detection and slot filling.

Main Results:

• Achieved 97.90% intent accuracy and 98.86% slot filling F1-score on the ATIS dataset.
• Attained 98.88% intent accuracy and 97.07% slot filling F1-score on the SNIPS dataset.
• Significantly outperformed baseline models on both datasets.

Conclusions:

• The proposed hybridized text representation effectively enhances joint intent detection and slot filling.
• The model demonstrates superior performance, advancing dialogue systems.
• Paves the way for more accurate and efficient NLU in real-world applications.