III: Medium: Generative Neural Information Retrieval Models

搜索引擎已经成为现代社会的主要信息来源。无论我们走到哪里，它们几乎把关于任何话题的海量知识都放在我们的指尖上。为了做到这一点，搜索引擎研究我们的搜索查询，并试图形成对我们首先要查找的内容的理解，当查询“Boston Tea Party Reason”、“IRS Form 1040”或“Piza Near Me”时。然后，将这种内部表示与数十亿个网页、书籍或新闻文章进行比较，试图为我们的搜索者的查询找到可能的最佳信息。该项目将以两种方式支持和创新这一进程。首先，它将通过对人类使用语言方式的洞察，而不仅仅是逐字比较查询和文档，来更准确地理解搜索查询。其次，使用这些改进的查询含义表示法，研究人员将开发一种从根本上不同的信息搜索方式。他们不是将我们的查询与每一个可能的匹配进行比较，而是让搜索引擎对查询提出一个理想化的响应，然后尝试找到与这个最佳答案最相似的网页。预期的结果将是更好的搜索结果和运行搜索引擎的机器更快的计算(这反过来可以减少电力需求和二氧化碳排放)。为了确保在该项目结束后仍能产生持久的影响，研究小组将积极联系搜索引擎公司的研究人员和工程师，以筹集更多资金，以促进该项目开发的技术的广泛采用。为了扩大对计算机科学的参与，使其超越传统上具有代表性的人口统计学(例如，就性别、种族或社会经济背景而言)，研究小组将在大学和中学水平的学生群体中举办一系列技术素养拓展活动。在这些活动中，研究人员将在来自不同背景的本科生的支持下，向学生参与者介绍有效的信息搜索工具和策略、该项目的研究目标以及大学水平的计算机科学教育。深度学习和表征学习为各种信息检索(IR)任务带来了可喜的改进。现有的神经IR模型使用从大量相关信息中学习的术语之间的语义相似性来估计信息需求(如查询或问题)与文档之间的匹配分数。与匹配分数的估计仅限于包含查询词的那些文档的经典IR模型不同，神经IR模型需要跟踪所有文档，或者改为重新排序从经典IR模型获得的排名靠前的文档。此外，由于神经IR模型通常基于术语含义的纯分布表示，它们缺乏对语言微妙之处的扎根理解，例如可分级术语。这个项目的目标是设计产生式信息检索模型，该模型通过分级术语的分布式表示来增强。为了实现这一目标，研究小组计划实现以下具体目标。(1)生成性IR模型：文档生成性模型可以有效地逼近给定查询的相关子空间中的表示，从而促进高效的全神经文档检索，而不是计算每个查询-文档对的匹配分数。研究人员将探索生成模型来近似相关文档的层次表示，并使用高效的最近邻算法来查找和检索集合中最合适的有机文档。(2)分级术语的分布式表示：通过考虑每个术语的全局上下文，可以解决分级术语通常是无形的含义。该项目将研究根据收集的数据估计的分级术语的假设价值范围和参照系的概率公式。(3)将分级术语表示纳入生成式检索系统：将分级术语的基础表示集成到生成式检索模型中，可以更好地理解和支持信息需求。该项目将研究这种对信息需求的影响，包括分级条款和不分级条款。该项目产生的预期成果是同行评议的科学出版物、建议模型的开源实现、预先训练的单词和短语嵌入、以TREC_VERA格式记录的检索运行、手动注释的主观蕴涵数据集，以及一套中学搜索素养教育材料。所有这些都将在知识共享CC0许可下在项目网站上共享。该项目由信息与智能系统部门的信息集成与信息学计划和既定的激励竞争研究计划(EPSCoR)共同资助。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Search engines have become modern society's main sources of information. They put vast amounts of knowledge about virtually any topic at our fingertips wherever we go. To do so, a search engine studies our search query and tries to form an understanding of what it is that we were looking for in the first place, when querying for "boston tea party reason", "IRS form 1040" or "pizza near me". This internal representation is then compared with billions of webpages, books, or news articles in an attempt to find the best possible information for our searcher's query. This project will support and innovate this process in two ways. First, it will develop a more accurate understanding of search queries using insights into the way that humans use language, rather than just comparing queries and documents word-by-word. Secondly, using these improved representations of query meaning, the researchers will develop a fundamentally different way of searching for information. Instead of comparing our query with every possible match, they let the search engine come up with an idealized response to the query and then try to find those webpages that are most similar to this optimal answer. The expected consequences will be better search results and faster computation for the machines running the search engines (that, in turn, can lead to reduced electricity demand and CO2 emissions). To ensure a lasting impact even after this project has concluded, the research team will actively reach out to researchers and engineers at search engine companies to raise facilitate widespread adoption of the technology developed in this project. To broaden participation in computer science beyond traditionally well-represented demographics (e.g., in terms of genders, ethnicities, or socio-economic backgrounds), the study team will host a range of technology literacy outreach events among student populations at the college and middle school level. During these events, the researchers, supported by undergraduates from diverse backgrounds, will inform the student participants about effective information search tools and strategies, the research goals of this project, and college-level computer science education in general.Deep and representation learning have brought promising improvements to various Information Retrieval (IR) tasks. Existing neural IR models estimate a matching score between the information need - such as a query or question - and the documents, using semantic similarities between terms, learned from a large set of relevance information. In contrast to classical IR models where the estimation of matching scores is constrained to only those documents containing the query terms, neural IR models need to trace over all documents, or instead re-rank the top-retrieved documents, obtained from a classical IR model. In addition, since neural IR models are often based on purely distributional representations of term meaning, they lack a grounded understanding of language subtleties such as for example gradable terms. The objective of this project is to design generative information retrieval models enhanced by distributed representations of gradable terms. To accomplish this, the research team plans the following concrete objectives. (1) Generative IR models: Instead of computing matching scores for each query-document pair, a document generative model can effectively approximate a representation in the relevance sub-space for a given query, facilitating efficient fully-neural document retrieval. The investigators will explore generative models to approximate hierarchical representations of relevant documents, and use efficient nearest-neighbor algorithms to find and retrieve the most suitable organic documents in the collection. (2) Distributed representations of gradable terms: The often intangible meaning of gradable terms can be resolved by considering the global context of each term. The project will study a probabilistic formulation of gradable terms based on their hypothetical value ranges and frames of reference, estimated from the collection. (3) Incorporation of gradable term representations into generative retrieval systems: The integration of grounded representations of gradable terms in the generative retrieval model will provide better understanding and support of information needs. The project will study this effect on information needs with and without gradable terms. The expected artifacts produced by the project are peer-reviewed scientific publications, open-source implementations of the proposed models, pre-trained word and phrase embeddings, logged retrieval runs in trec_eval format, a manually annotated Subjective Entailment dataset, and a suite of middle school search literacy education materials. All of these will be shared on the project website under Creative Commons CC0 License.This project is jointly funded by the Information Integration & Informatics Program in the Division of Information & Intelligent Systems, and the Established Program to Stimulate Competitive Research (EPSCoR).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.