HeterMPC: A Heterogeneous Graph Neural Network for Response Generation in Multi-Party Conversations

Recently, various response generation models for two-party conversations have\nachieved impressive improvements, but less effort has been paid to multi-party\nconversations (MPCs) which are more practical and complicated. Compared with a\ntwo-party conversation where a dialogue context is a sequence of utterances,\nbuilding a response generation model for MPCs is more challenging, since there\nexist complicated context structures and the generated responses heavily rely\non both interlocutors (i.e., speaker and addressee) and history utterances. To\naddress these challenges, we present HeterMPC, a heterogeneous graph-based\nneural network for response generation in MPCs which models the semantics of\nutterances and interlocutors simultaneously with two types of nodes in a graph.\nBesides, we also design six types of meta relations with\nnode-edge-type-dependent parameters to characterize the heterogeneous\ninteractions within the graph. Through multi-hop updating, HeterMPC can\nadequately utilize the structural knowledge of conversations for response\ngeneration. Experimental results on the Ubuntu Internet Relay Chat (IRC)\nchannel benchmark show that HeterMPC outperforms various baseline models for\nresponse generation in MPCs.\n