Główne pojęcia
MP2D proposes a framework for generating dialogues with natural topic transitions by leveraging knowledge graphs.
Streszczenie
MP2D introduces a novel approach to automatically generate dialogues with smooth topic shifts. By utilizing knowledge graphs, the framework maps the flow of topics within a dialogue, creating high-quality datasets for training models. Experimental results demonstrate the efficacy of MP2D in generating dialogues with natural topic transitions and showcase its potential in various topic shift dialogue tasks.
Statystyki
MP2D retrieves passages using entities in a knowledge graph.
GPT-3.5 outperforms other question generators across all metrics.
Models trained on MP2D-generated datasets show superior performance in topic segmentation and detection tasks.