School of Computer Science and Engineering, Beijing Key Laboratory of Network Technology, Beihang University, Beijing 100191, China.
School of Cyber Science and Technology, Beihang University, Beijing 100191, China.
Sensors (Basel). 2020 May 28;20(11):3057. doi: 10.3390/s20113057.
Modern retrieval systems tend to deteriorate because of their large output of useless and even misleading information, especially for complex search requests on a large scale. Complex information retrieval (IR) tasks requiring multi-hop reasoning need to fuse multiple scattered text across two or more documents. However, there are two challenges for multi-hop retrieval. To be specific, the first challenge is that since some important supporting facts have little lexical or semantic relationship with the retrieval query, the retriever often omits them; the second challenge is that once a retriever chooses misinformation related to the query as the entities of cognitive graphs, the retriever will fail. In this study, in order to improve the performance of retrievers in complex tasks, an intelligent sensor technique was proposed based on a sub-scope with cognitive reasoning (2SCR-IR), a novel method of retrieving reasoning paths over the cognitive graph to provide users with verified multi-hop reasoning chains. Inspired by the users' process of step-by-step searching online, 2SCR-IR includes a dynamic fusion layer that starts from the entities mentioned in the given query, explores the cognitive graph dynamically built from the query and contexts, gradually finds relevant supporting entities mentioned in the given documents, and verifies the rationality of the retrieval facts. Our experimental results show that 2SCR-IR achieves competitive results on the HotpotQA full wiki and distractor settings, and outperforms the previous state-of-the-art methods by a more than two points absolute gain on the full wiki setting.
现代检索系统往往会因为输出大量无用甚至误导性的信息而恶化，尤其是对于大规模的复杂搜索请求。需要进行多跳推理的复杂信息检索 (IR) 任务需要融合来自两个或更多文档的多个分散的文本。然而，多跳检索面临两个挑战。具体来说，第一个挑战是，由于一些重要的支持事实与检索查询几乎没有词汇或语义关系，检索器经常忽略它们；第二个挑战是，一旦检索器选择与查询相关的错误信息作为认知图的实体，检索器就会失败。在这项研究中，为了提高复杂任务中检索器的性能，提出了一种基于具有认知推理的子范围的智能传感器技术 (2SCR-IR)，这是一种在认知图上检索推理路径的新方法，为用户提供经过验证的多跳推理链。受用户在线逐步搜索过程的启发，2SCR-IR 包含一个动态融合层，该层从给定查询中提到的实体开始，从查询和上下文中动态构建认知图，逐步找到给定文档中提到的相关支持实体，并验证检索事实的合理性。我们的实验结果表明，2SCR-IR 在 HotpotQA 全维基和干扰设置上取得了有竞争力的结果，在全维基设置上比之前的最先进方法高出超过两个百分点的绝对增益。
