III: Small: Searching for Answers through Iterative Feedback

在当前的网络搜索引擎中，对查询的响应通常是包含排序结果的一系列页面(搜索引擎结果页面或SERP)。移动搜索的使用越来越多，这使得有限的可用显示空间得到了极大的利用。同样，基于语音的搜索，即通过语音识别和语音生成来完成问题和答案，正变得越来越普遍，这也对系统和用户之间的交互带宽造成了限制。在这些情况下，能够对广泛的问题提供更准确的答案，而不是按顺序显示结果，变得至关重要。如果搜索系统可以返回可能答案的排序列表而不是文档，并且搜索环境可能限制用户系统带宽，这将导致以下重要的研究问题，该问题是该提议的焦点--呈现排序答案列表并与其交互的最有效方式是什么，其中的目标是尽可能快地识别一个或多个令人满意的答案。这个项目将致力于四个研究任务：(A)开发和评估迭代的答案相关反馈模型；(B)开发和评估答案的交互式摘要技术；(C)开发和评估更细粒度的答案反馈方法；(D)开发和评估基于对话的答案检索模型。这个项目将是第一个研究与答案排序列表交互的方法和模型的项目。许多研究人员正在为事实类问题回答任务开发神经模型，但这项努力是少数几个研究在文件段落中找到非事实答案的问题的努力之一。为这一复杂任务开发神经模型所获得的经验为本提案中描述的独特任务和方法提供了背景，这些任务和方法解决了关键但以前被忽视的问题，即我们如何有效利用答案的排序列表与用户交互，并改进神经答案检索模型的结果。该项目的后半部分将解决在搜索中使用对话模式的问题，这也变得越来越重要，但尚未被研究。

In current web search engines, the response to a query is typically a series of pages that contain ranked results (search engine result pages or SERPs). The increasing use of mobile search places a premium on the use of the limited display space that is available. Similarly, voice-based search, where both questions and answers are done by voice recognition and speech generation, is becoming more common and also creates a limitation on the interaction bandwidth between the system and the user. In these situations, the ability to deliver more precise answers to a broad range of questions, rather than a ranked display of results, becomes critical. If a search system can return a ranked list of possible answers instead of documents, and a search environment may limit the user-system bandwidth, this leads to the following important research question that is the focus of this proposal -- what is the most effective way to present and interact with a ranked list of answers, where the goal is to identify one or more satisfactory answers as quickly as possible. Understanding this problem and discovering solutions to it will have a large impact on the future development of search engines.This project will work on four research tasks: (a) develop and evaluate iterative relevance feedback models for answers; (b) develop and evaluate interactive summarization techniques for answers; (c) develop and evaluate finer-grained feedback approaches for answers; (d) develop and evaluate a conversation-based model for answer retrieval. This project will be the first to study methods and models for interacting with ranked lists of answers. Many researchers are developing neural models for the factoid question-answering task, but this effort is one of just a few looking at the problem of finding non-factoid answers in passages of documents. The experience gained from developing neural models for this complex task provides the background for the unique tasks and approaches described in this proposal, which address the key, but previously ignored, issue of how we make effective use of ranked lists of answers to interact with users and improve the results from neural answer retrieval models. The later part of the project will address the use of conversational models in search, which is also becoming increasingly important but has not yet been studied.