报告题目：New directions in Conversational Question Answering
报告人：德国马克思普朗克信息学研究所Rishiraj Saha Roy高级研究员
报告时间：2023年03月16日（周四）下午16:00
报告地点：腾讯会议（会议号：819-681-686）
主持人：贾真博士

报告摘要：Conversational question answering (ConvQA) around factual topics is becoming a popular mechanism of interaction with personal assistants. In this talk, I will discuss two research directions of prime importance in ConvQA - learning from user feedback, and operating over heterogeneous sources. For user feedback, we deviate from the standard paradigm of gold-standard QA pairs that are tedious to obtain in real-life. Instead, we tap into from the more natural inputs of question reformulations in a reinforcement learning setup. On the perspective of heterogeneous input sources, we propose converting incomplete conversational utterances into self-contained structured representations that are more amenable to matching over multiple sources of knowledge. The conversion of partial questions to intent-explicit SRs is done via fine-tuned transformer models. The talk will end with promising directions of future work in this paradigm.
报告人简介：Rishiraj Saha Roy is a Senior Researcher at the Max Planck Institute for Informatics (MPII), Saarbruecken, Germany, leading the research group on Question Answering, that focuses on building robust and interpretable solutions for answering natural language questions over structured and unstructured sources. Some of his most important work has been published in international conferences like SIGIR, WWW, WSDM, EMNLP, CIKM, etc.