SOLID: Self-seeding and Multi-intent Self-instructing LLMs for Generating Intent-aware Information-Seeking Dialogs
Arian Askari, Roxana Petcu, Chuan Meng, Mohammad Aliannejadi, Amin Abolghasemi, Evangelos Kanoulas, Suzan Verberne · 2025
Intent prediction in information-seeking dialogs is challenging and requires a substantial amount of data with human-labeled intents for effective model training.While Large Language Models (LLMs) have demonstrated effectiveness in generating synthetic data, existing methods typically rely on human feedback and are tailored to structured, task-oriented intents.In this paper, we leverage LLMs for zero-shot generation of large-scale, opendomain, intent-aware information-seeking dialogs to serve as training data for intent prediction models.We introduce SOLID, a method that generates dialogs turn by turn using novel self-seeding and multi-intent self-instructing strategies.Additionally, we propose SOLID-RL, a finetuned version that generates an entire dialog in one step using data created with SOLID.SOLID and SOLID-RL are each used to generate over 300k intent-aware dialogs, significantly surpassing the size of existing datasets.Experiments show that intent prediction models trained on sampled dialogs generated by SOLID and SOLID-RL outperform those trained solely on human-generated dialogs.Our findings demonstrate the potential of LLMs to expand training datasets, as they provide valuable resources for conversational agents across multiple tasks.Our self-seeding and self-instructing approaches are adaptable to various conversational data types and languages with minimal modifications.