Operationalizing the Logical Uncertainty Principle in a Language Modelling Framework for Context-based Information Retrieval

基于上下文的信息检索（IR）是近年来基于上下文的信息检索技术的革命性兴起。IR研究正在向以上下文为中心的方法转变，使人们能够从他/她的检索上下文中进行推断。我坚信上下文驱动的推理是构建更多用户和上下文敏感的IR系统的关键。在与IR相关的应用程序中（例如，检索给定查询的相关文档，用相关术语扩展原始查询，在问答中确定正确答案，或在跨语言检索中确定适当的翻译），由于信息转换和上下文依赖关系，通常会嵌入某些形式的推理。举例来说，给定一个Java查询，IR系统可能返回关于编程的文档和关于Merapi（爪哇岛中部的一座火山）的文档，因为它们都包含术语Java。如果检索上下文是计算机，则有关编程的文档是相关的。然而，对于火山学家来说，关于默拉皮火山的文件更有可能是相关的。上述不确定性来自于从Java和计算机到编程的信息流，或者从Java和火山到默拉皮的信息流，这取决于检索上下文。有必要了解当前的IR技术，包括当前的搜索引擎，如b谷歌，并不能满足刚才给出的场景。本研究的动机是以下基本问题：我们能否使推理过程在IR系统中明确，从而获得推理能力，根据他/她的检索上下文选择真正相关的信息项？逻辑不确定性原理将红外视为一个合理的逻辑推理过程，从而为基于上下文的红外提供了潜在的重要理论基础。然而，由于难以获得上下文和领域知识以及难以大规模实现符号逻辑模型，其可操作性一直是一个问题。语言技术的最新进展为逻辑不确定性原理在实际环境中的实现打开了大门。最近，语言建模框架已经被开发出来，通过平滑机制集成不同类型的术语关系。语言建模方法提供了一个坚实的理论背景，产生了有希望的实验结果（与最好的红外系统相媲美），并且计算效率也很高。基于我在这个方向上的现有工作，我将研究逻辑不确定性原则在语言建模框架中的操作化，以促进有效的上下文依赖推理。我将进行理论研究，原型系统开发，以及大规模数据集的实验评估。我相信，将逻辑推理和语言建模的优势结合起来，作为新一代IR基础设施，可以产生更智能和上下文敏感的IR系统，同时在计算上易于处理。

The pressing need to deal with information overload has brought about the recent revolutionary emergence of context-based information retrieval (IR). IR research is experiencing a shift to context-centric approaches enabling one to infer from his/her retrieval context. I strongly believe that context-driven reasoning is the key in building more user and context-sensitive IR systems. In applications related to IR (for example, retrieving relevant documents to a given query, expanding the original query with related terms, determining the correct answer in Question-Answering, or determining an appropriate translation in cross-language retrieval), some forms of reasoning are often embedded as a result of information transformation and context dependency. As an illustration, given a query Java , an IR system may return documents about programming and documents about Merapi (a volcano in central Java island), as they all contain the term java . If the retrieval context is computer , documents about programming are relevant. However, for a volcanologist, documents about Merapi are more likely to be relevant. The above uncertainty arises from the flow of information from Java and computer to programming , or from Java and volcano to Merapi , depending on the retrieval context. It is essential to understand that current IR technology, including current search engines such as Google, does not cater for scenarios like that just given.This research is motivated by the following fundamental question: can we make the reasoning process explicit in an IR system which can in turn gain a capability of reasoning to select truly relevant information items depending on his/her retrieval context? The Logical Uncertainty Principle views IR as a plausible logical inference process and thus provides a potentially significant theoretical foundation for context-based IR. Nevertheless, its operationalization has long been a problem, due to the difficulty with obtaining the contextual and domain knowledge as well as implementing the symbolic logical models on a large scale. Recent advances in language technologies open the door to realizing the Logical Uncertainty Principle in a practical setting. Recently, language modelling frameworks have been developed for IR to integrate different types of term relationships via a smoothing mechanism. The language modelling approach provides a solid theoretical setting, produces promising experimental results (comparable to the best IR systems), and is also computationally efficient. Based on my existing work in this direction, I will investigate the operationalization of the Logical Uncertainty Principle in a language modelling framework to facilitate effective context-dependent reasoning. I will conduct theoretical research, prototyping system development, and experimental evaluation with large-scale datasets. It is my belief that the combination of the strengths of logical inference and language modelling as a new generation IR infrastructure can lead to more intelligent and context-sensitive but at the same time computationally tractable IR systems.