Extracting Triples from Dialogues for Conversational Social Agents

Understanding of communication is essential to create controllable and transparent agents for Hybrid Intelligence collaboration. In this paper, we describe a number of Natural Language Understanding models that extract explicit symbolic triples from social conversation. Triple extraction has mostly been developed for Knowledge Base Completion using Wikipedia data for training and testing. However, social conversation is a very different genre. Phenomena such as co-reference, ellipsis, coordination, and implicit and explicit negation or confirmation are more prominent in conversation. We therefore fill this gap by releasing data sets for triple extraction from social conversation. We created five new triple extraction models and tested seven models on our data. The highest precision was achieved by a generative model, Qwen2.5: 54.02 for complete triples and 73.18 for triple elements tested on single utterances. However, scores for conversational triples that span multiple turns are significantly lower, showing that extracting knowledge from true conversational data is challenging.