Abstract
We study multi-turn response generation for open-domain dialogues. The existing state-of-the-art addresses the problem with deep neural architectures. While these models improved response quality, their complexity also hinders the application of the models in real systems. In this work, we pursue a model that has a simple structure yet can effectively leverage conversation contexts for response generation. To this end, we propose four auxiliary tasks including word order recovery, utterance order recovery, masked word recovery, and masked utterance recovery, and optimize the objectives of these tasks together with maximizing the likelihood of generation. By this means, the auxiliary tasks that relate to context understanding can guide the learning of the generation model to achieve a better local optimum. Empirical studies with three benchmarks indicate that our model can significantly outperform state-of-the-art generation models in terms of response quality on both automatic evaluation and human judgment, and at the same time enjoys a much faster decoding process.
Introduction
Multi-turn response generation in open-domain dialogues requires models to understand complex, hierarchical contexts where words form utterances and utterances form a conversation. While state-of-the-art (SOTA) models have improved response quality using deep neural architectures like ReCoSa, their complexity often hinders real-world deployment due to high parameter counts and slow decoding. This work explores whether a simpler model architecture can achieve comparable or superior results by shifting the burden of context understanding from model depth to the training objective.
Model Architecture
The proposed model utilizes a lightweight, single-layer Transformer-based encoder-decoder as its backbone. To enhance its ability to capture dialogue structure, the model is trained using a joint learning objective that combines standard Maximum Likelihood Estimation (MLE) with four self-supervised auxiliary tasks.
The encoder processes conversation history through word, position, and segment embeddings followed by a Transformer layer. The decoder then generates responses using context-response attention. The "Auxiliary Tasks" module branches from the encoder's output, allowing these tasks to guide the learning of rich context representations without increasing the model's inference-time complexity.
Auxiliary Tasks
The core innovation lies in four auxiliary tasks designed to teach the model different aspects of dialogue structure:
- Word Order Recovery (WOR): The model reconstructs the original word order from a shuffled utterance, learning sequential dependencies.
- Masked Word Recovery (MWR): Similar to BERT, the model predicts masked words (15%) based on surrounding context.
- Masked Utterance Recovery (MUR): The model predicts the words of an entirely masked utterance, forcing it to rely on the broader conversation flow.
- Utterance Order Recovery (UOR): The model predicts the correct chronological order of shuffled utterances, capturing high-level dialogue logic.
Experimental Results
The model was evaluated on three benchmarks: DailyDialog, PERSONA-CHAT, and the Ubuntu Dialogue Corpus (UDC).
Performance and Efficiency
The simple 1-layer model significantly outperformed complex SOTA baselines (e.g., ReCoSa, HRAN) in both automatic metrics (Perplexity, BLEU, Diversity) and human evaluations. Notably, the model achieved a "depth equivalence," where a single layer trained with auxiliary tasks matched the performance of much deeper models (up to 7 layers on UDC) trained with standard MLE.
Crucially, the model is 2x faster at decoding than ReCoSa and maintains a much smaller footprint (e.g., 14.4M vs. 73.8M parameters on DailyDialog), making it highly suitable for real-time applications.
Ablation Study
Ablation tests confirmed that each auxiliary task contributes uniquely to performance. For instance, UOR was particularly effective for DailyDialog's narrative flow, while word-level tasks (WOR/MWR) were vital for PERSONA-CHAT, where factual consistency is key.
Conclusion
This research demonstrates that clever training objectives can compensate for architectural simplicity in dialogue systems. By using auxiliary tasks to "teach" context understanding, we can build models that are both more effective and significantly more efficient than their deeper counterparts. Future work will investigate the scalability of these tasks to even more diverse dialogue scenarios and their impact on out-of-distribution generalization.
Summary updated