Abstract
Hierarchical neural networks are often used to model inherent structures within dialogues. For goal-oriented dialogues, these models miss a mechanism adhering to the goals and neglect the distinct conversational patterns between two interlocutors. In this work, we propose Goal-Embedded Dual Hierarchical Attentional Encoder-Decoder (G-DuHA) able to center around goals and capture interlocutor-level disparity while modeling goal-oriented dialogues. Experiments on dialogue generation, response generation, and human evaluations demonstrate that the proposed model successfully generates higher-quality, more diverse and goal-centric dialogues. Moreover, we apply data augmentation via goal-oriented dialogue generation for task-oriented dialog systems with better performance achieved.
Introduction
Building a chatbot that can handle a specific task, such as booking a train or finding a hotel, is a complex challenge. Standard hierarchical models, like the Hierarchical Recurrent Encoder-Decoder (HRED), are effective at modeling turn-by-turn conversation flow but often fail in task-oriented settings due to two primary issues:
- Goal Drift: Without a mechanism to maintain the objective, models often deviate into irrelevant topics, switching from booking a taxi to discussing museums.
- Role Confusion: Users and agents have distinct conversational roles (asking vs. assisting). Standard models often use shared parameters, leading to "role confusion" where the agent adopts user-like patterns or vice versa.
The proposed Goal-Embedded Dual Hierarchical Attentional Encoder-Decoder (G-DuHA) addresses these gaps by centering the model around specific goals and acknowledging the distinct patterns of the two interlocutors.
Methods
The G-DuHA model introduces three architectural innovations to ensure task adherence and role distinction:
- Goal Embedding Module: This module takes the set of goals (e.g., Domain: Train, Request: Price) and processes them through a Feed-Forward Network (FFN). The resulting "goal context" vector initializes the hidden states of the dialogue's recurrent neural networks (RNNs), serving as a persistent "North Star."
- Dual Hierarchical Architecture: G-DuHA employs two separate hierarchical structures—one for the user and one for the agent. This captures the "interlocutor-level disparity" inherent in task-oriented dialogues.
- Multi-Level Attention: The model uses attention at both the word level (identifying key words in a sentence) and the dialogue level (weighing the importance of previous turns).
Experimental Setup
The model was trained and evaluated using the MultiWOZ dataset, which contains over 8,000 human-human dialogues across multiple domains. For data augmentation tests, the WoZ restaurant reservation dataset was used to evaluate if synthetic dialogues could improve the performance of the GLAD Dialogue State Tracker.
Key Results
G-DuHA was compared against several baselines, including standard HRED and goal-augmented Language Models (LM+G).
- Quality and Diversity: G-DuHA achieved a BLEU score of 7.48, outperforming HRED's 5.40. Notably, utterance-level diversity (D-U) jumped to 25.98, compared to HRED's 3.94, indicating that role-specific modeling significantly reduces repetitiveness.
- Goal Adherence: Human evaluations showed that G-DuHA maintained "Goal Focus" 82.33% of the time when compared directly to HRED.
- Data Augmentation: Using G-DuHA to generate synthetic training data for the GLAD tracker improved its "Turn Request" accuracy from 97.11% to 97.59%, a statistically significant gain.
Conclusion
The G-DuHA model demonstrates that explicitly embedding goals and separating interlocutor roles are critical for high-quality task-oriented dialogue generation. By providing a structural bias toward the objective, the model avoids goal drift and generates more diverse, human-like responses. Future work may involve visualizing goal embeddings to further understand task interpretation and applying this dual-hierarchical approach to broader conversational AI domains.
Summary updated