Speaker Profiling in Multi-party Conversations

In a conversation, individual speakers respond uniquely. Consequently, a `one size fits all' technique is not the best way for a dialog agent to generate responses. While many studies design personalized dialog agents with the help of persona information of speakers, all of them assume that speaker persona is supplied beforehand. However, this assumption does not hold for all conversational agents. This paper attempts to address this gap by exploring the task -- Speaker Profiling in Conversations (SPC). The aim of SPC is to produce a summary of the persona characteristics for individual speakers present in a dialog. We divide this task into two subtasks -- persona discovery and persona-type identification. Given a dialog, the former subtask aims to find all its utterances that carry persona information. The latter subtask focuses on evaluating these utterances to identify the type of persona information present in them. We present SPICE, a novel dataset that is specifically curated with labels to tackle the task of SPC. We show the performance of various baselines on SPICE and benchmark it using SPOT, a neural model based on GRU and Transformer. SPOT shows an improvement of ~7% for persona discovery and ~3% for persona-type identification over the best baseline. We further present a thorough analysis of SPOT to diagnose individual modules and their limitations, both quantitatively and qualitatively.