A Bi-model based RNN Semantic Frame Parsing Model for Intent Detection and Slot Filling (1812.10235v1)

Abstract: Intent detection and slot filling are two main tasks for building a spoken language understanding(SLU) system. Multiple deep learning based models have demonstrated good results on these tasks . The most effective algorithms are based on the structures of sequence to sequence models (or "encoder-decoder" models), and generate the intents and semantic tags either using separate models or a joint model. Most of the previous studies, however, either treat the intent detection and slot filling as two separate parallel tasks, or use a sequence to sequence model to generate both semantic tags and intent. Most of these approaches use one (joint) NN based model (including encoder-decoder structure) to model two tasks, hence may not fully take advantage of the cross-impact between them. In this paper, new Bi-model based RNN semantic frame parsing network structures are designed to perform the intent detection and slot filling tasks jointly, by considering their cross-impact to each other using two correlated bidirectional LSTMs (BLSTM). Our Bi-model structure with a decoder achieves state-of-the-art result on the benchmark ATIS data, with about 0.5$\%$ intent accuracy improvement and 0.9 $\%$ slot filling improvement.

• The paper introduces a bi-model RNN that leverages shared BLSTM hidden states to capture the interplay between intent detection and slot filling.
• Methodology employs dual network structures with asynchronous training and distinct cost functions to improve accuracy on both tasks.
• Empirical evaluations show state-of-the-art boosts on the ATIS and multi-domain datasets, highlighting its real-world applicability in SLU systems.

Analysis of a Bi-model based RNN Semantic Frame Parsing Model for Intent Detection and Slot Filling

The paper under review, titled "A Bi-model based RNN Semantic Frame Parsing Model for Intent Detection and Slot Filling," introduces innovative RNN structures designed to enhance the performance of spoken language understanding (SLU) systems. The research specifically targets two central tasks in SLU: intent detection and slot filling, typically considered independent yet interrelated challenges.

Detailed Contributions and Methodology

The authors propose a Bi-model approach utilizing Recurrent Neural Networks (RNNs), more specifically, bi-directional Long Short-Term Memory (BLSTM) networks. They develop two network structures: one incorporating an LSTM-based decoder and the other without it. The core novelty lies in modeling the cross-impact between intent detection and slot filling by allowing two task-networks to share hidden states, thus enabling mutual information exchange and enhancing performance for both tasks.

In conventional SLU systems, separate models or joint sequence-to-sequence (S2S) models are employed, which either handle the tasks in parallel or leverage joint architectures that do not adequately capture the interdependencies between the two tasks. This paper posits that by designing correlated models with asynchronous training based on distinct cost functions—cross entropy for both intent detection and slot filling—a more comprehensive understanding of the input can be achieved, leading to improved accuracy.

Experimental Results

The efficacy of the proposed models is evaluated on the well-established ATIS dataset and a proprietary multi-domain dataset encompassing food, home, and movie domains. The empirical results are notable. The model utilizing a decoder achieved state-of-the-art performance on the ATIS dataset, yielding a 0.5% improvement in intent accuracy and a 0.9% improvement in slot filling F1 scores. Such results underscore the relative improvements over baseline models, fundamentally reducing errors in test datasets by substantial margins.

On the multi-domain dataset, comparisons with the current best model, the attention-based BiRNN, further affirmed the superiority of the Bi-model approach. The proposed model consistently demonstrated enhancements in both F1 scores and intent detection accuracy across different domains, emphasizing its adaptability and robustness in varying contexts.

Implications and Future Directions

This research significantly impacts the practical deployment of SLU systems in real-world scenarios. By effectively capturing the interplay between intent and semantic tagging, these models can improve the accuracy of natural language understanding applications, such as virtual assistants, customer service bots, and interactive voice response systems.

Looking ahead, several avenues merit exploration. Extending this approach to larger, multilingual datasets could validate and extend the generalizability of the model. Additionally, integrating and leveraging other deep learning paradigms, such as transformer architectures, might further capitalize on the cross-task synergy demonstrated herein. Exploring the model's efficiency in terms of computational resource requirements also presents an opportunity for optimization, potentially broadening the accessibility of sophisticated SLU models.

In conclusion, the Bi-model based RNN framework presents a compelling advancement in semantic frame parsing, demonstrating clear improvements over existing methodologies and offering a blueprint for future investigations into joint task modeling within SLU systems.