Synthetic dialogues generated by large language models (LLMs) differ from real dialogues in linguistic attributes such as naturalness or sentence completeness.
To bridge this gap, we propose AlignedAug, a framework for realistic dialogue augmentation....
Synthetic dialogues generated by large language models (LLMs) differ from real dialogues in linguistic attributes such as naturalness or sentence completeness.
To bridge this gap, we propose AlignedAug, a framework for realistic dialogue augmentation. AlignedAug consists of three stages:
(1) Cognition-aware Dialogue Generation, which generates utterances using a LLM-based model; (2) Chat Style Refinement, which simulates informal, chat-like responses by randomly deleting words except for subjects, verbs, and negations; and (3) Statistical Selection, which selects dialogues whose naturalness and coherence scores are aligned with scores of real dialogues.
Experimental results show that the Chat Style Refinement and Statistical Selection stages significantly reduce the gap between synthetic and real dialogues. On the CollabChat dataset, the $S_{KS}$ score decreases from 0.95 to 0.35. AlignedAug also outperforms existing LLM-based dialogue augmentation methods on classification and response selection tasks. On the CollabChat dataset, classification accuracy improves from 0.65 to 0.69, and R@5 in the response selection task increases from 0.77 to 0.84.
These findings demonstrate that AlignedAug provides synthetic data that not only augments dialogues which align real dialogues more closely but also improves the performance of models trained on aligned dialogues.