Context Modeling for IQA: The Role of Tasks and Entities

In a realistic Interactive Question Answer-ing (IQA) setting, users frequently ask follow-up questions. By modeling how the questions ’ focus evolves in IQA dialogues, we want to describe what makes a partic-ular follow-up question salient. We intro-duce a new focus model, and describe an implementation of an IQA system that we use for exploring our theory. To learn prop-erties of salient focus transitions from data, we use logistic regression models that we validate on the basis of predicted answer correctness. 1 Questions within a Context Question Answering (QA) systems have reached a high level of performance within the scenario orig-inally described in the TREC competitions, and are ready to tackle new challenges as shown by the new tracks proposed in recent instantiations (Voorhees, 2004). To answer these challenges, at-tention is moving towards adding semantic infor-mation at different levels. Our work is about con-text modeling for Interactive Question Answering (IQA) systems. Our research hypothesis is that a) knowledge about the dialogue history, and b) lexi-cal knowledge about semantic arguments improve an IQA system’s ability to answer follow-up ques-tions. In this paper we use logistic regression mod-eling to verify our claims and evaluate how the per-formance of our Q→A mapping algorithm varies based on whether such knowledge is taken into ac-count