Conditional Joint Model for Spoken Dialogue System

Changliang Li, Zhao Yan, Dong YuLecture Notes in Computer Science 2019 · DOI · OpenAlex

Introduction

Traditional Spoken Dialogue Systems (SDS) typically rely on a pipeline of independent components: Spoken Language Understanding (SLU) to identify user intent and extract slots, followed by Dialogue Management (DM) to predict the next system action. While recent research has explored joint learning to train these components together, many existing models simply aggregate individual losses without facilitating deep information sharing between tasks.

This paper introduces a Conditional Joint Model (Con-Joint) that bridges this gap. By treating the user's predicted intent as a primary guiding condition for both slot filling and system action prediction, the model leverages the inherent correlations between these tasks to achieve superior performance on real-world dialogue benchmarks.

Methodology

The proposed hierarchical architecture treats user intent as the central context for all subsequent dialogue tasks. The model operates through several integrated layers:

  • Shared Feature Layer: A bidirectional Long Short-Term Memory (Bi-LSTM) network processes the input utterance to generate a base semantic representation for each word.
  • Intent-as-Condition: The model first predicts the user's intent, which is then converted into a representative vector ($v^{int}$).
  • Conditional Slot Filling: Rather than labeling slots based solely on word embeddings, the model utilizes the intent vector as an attention weight. This mechanism allows the model to focus on sentence components most relevant to the specific predicted intent.
  • Conditional Action Prediction (SAP): To determine the system's next response, the model analyzes a window of previous dialogue turns, combining sentence-level features with the predicted intent vector. This data is processed by a top-level Bi-LSTM to predict final system actions.
The architecture of the SLU component within the Con-Joint model, illustrating how intent labels and slot tags are derived from shared embeddings.

By employing a united loss function, the model optimizes intent detection, slot filling, and action prediction simultaneously, encouraging the hidden layers to learn representations that benefit the entire dialogue flow.

Experimental Setup

The model was evaluated using the DSTC4 (Dialog State Tracking Challenge 4) corpus, which consists of real-world Skype calls between tour guides and tourists in Singapore. The dataset covers five domains: accommodation, attraction, food, shopping, and transportation. It is particularly challenging due to its origin in natural, unscripted human speech, featuring 5,648 training utterances with 87 slot labels and 68 intent types.

Key Results and Insights

The "Con-Joint" model demonstrated significant improvements over standard Conditional Random Fields (CRFs), Support Vector Machines (SVMs), and basic joint learning models across all primary metrics.

Comparative performance metrics for Slot Filling, Intent Detection, and overall SLU accuracy across different model architectures.
  • Slot Filling: The model achieved an F1-score of 49.32, representing a 3.17% improvement over the previous best Bi-LSTM baseline.
  • Intent Detection: Reached an F1-score of 49.81, outperforming independent methods.
  • System Action Prediction (SAP): This task saw the most substantial gain, with the model achieving an F1-score of 33.69, compared to the 31.15 achieved by the SVM baseline. Notably, the model showed a 10.05% improvement in recall for system actions, indicating a robust ability to identify correct actions in complex dialogues.

Core Insights

  1. Intent as a Bridge: The primary innovation is recognizing intent as the most informative signal. Making slot filling and action prediction conditional on intent mimics human cognitive processes, where understanding the goal precedes processing the details.
  2. Implicit Knowledge Sharing: Joint training allows the lower-level Bi-LSTM to learn features that are simultaneously useful for immediate understanding and future dialogue prediction.
  3. Multi-Label Handling: By using sigmoid functions instead of softmax for intent and action prediction, the model effectively handles utterances where users express multiple intents or the system must perform multiple simultaneous actions.

Conclusion and Future Work

While the Con-Joint model excels in F1-score and recall, it faces a trade-off in frame-level accuracy due to the high complexity of the DSTC4 dataset, where some turns contain up to 20 different action labels. Future research will focus on creating more explicit three-way feedback loops between slots, intents, and actions. Additionally, the authors plan to expand the framework to include Natural Language Generation (NLG), moving toward a fully end-to-end spoken dialogue system.

Summary updated

BibTeX
@inproceedings{li2019conditional,
  author = {Changliang Li and Zhao Yan and Dong Yu},
  title = {{Conditional Joint Model for Spoken Dialogue System}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2019},
  doi = {10.1007/978-3-030-23407-2\_3},
}
Figure