Attention-Based Joint Learning for Intent Detection and Slot Filling Using Bidirectional Long Short-Term Memory and Convolutional Neural Networks (AJLISBC)

Effective natural language understanding is crucial for dialogue systems, requiring precise intent detection and slot filling to facilitate interactions. Traditionally, these subtasks have been addressed separately, but their interconnection suggests that joint solutions yield better results. Recent neural network-based approaches have shown significant performance in joint intent detection and slot filling tasks. The two primary neural network structures used are recurrent neural networks (RNNs) and convolutional neural networks (CNNs). RNNs capture long-term dependencies and store previous information semantics in a fixed-size vector, but their ability to extract global semantics is limited. CNNs can capture n-gram features using convolutional filters, but their performance is constrained by filter width. To leverage the strengths and mitigate the weaknesses of both networks, this paper proposes an attention-based joint learning classification for intent detection and slot filling using BiLSTM and CNNs (AJLISBC). The BiLSTM encodes input sequences in both forward and backward directions, producing high-dimensional representations. It applies scalar and vectorial attention to obtain multichannel representations, with scalar attention calculating word-level importance and vectorial attention assessing feature-level importance. For classification, AJLISBC employs a CNN structure to capture word relations in the representations generated by the attention mechanism, effectively extracting n-gram features. Experimental results on the benchmark Airline Travel Information System (ATIS) dataset demonstrate that AJLISBC outperforms state-of-the-art methods.