基于文本知识的多跳推理综述

Survey of text knowledge-based multi-hop reasoning

基于文本知识的多跳推理是一种从大规模文本知识库中选择和组合相关知识,以回答复杂问题或验证假设的过程。这种能力可以显著提高自然语言处理系统的实用性,因此基于文本知识的多跳推理已经成为自然语言处理领域的研究热点。然而,该任务面临两大挑战:1)如何从海量语料库中准确地选择推理所需的知识;2)如何通过多步语义合成进行有效推理。针对这些挑战,该领域出现了许多优秀的工作,但是尚缺乏对该领域的综述。因此对该领域的相关工作进行梳理和总结。首先给出多跳推理以及基于文本知识的多跳推理的定义;其次,分析、比较了当前流行和高质量的基于文本知识的多跳推理数据集;再次,对已有的知识选择方法进行分类,并分析各类方法的优缺点;继次,对已有的推理方法进行分类讨论;最后,探讨了该领域未来可能的发展方向。

Text-based multi-hop reasoning is a process of selecting and combining relevant knowledge from large-scale text knowledge bases to answer complex questions or validate hypotheses.This capability can significantly enhance the practicality of natural language processing systems,so that text-based multi-hop reasoning becomes a research hot point in the field of natural language processing.However,this task faces two major challenges:1)how to accurately select the knowledge required for reasoning from massive corpora;2)how to perform effective reasoning through multi-step semantic synthesis.In response to these challenges,many outstanding works have emerged in this field,but a comprehensive review is still lacking.Therefore,relevant work in this field was compiled and summarized.Firstly,the definitions of multi-hop reasoning and text-based multi-hop reasoning were given.Secondly,the current popular and high-quality text-based multi-hop reasoning datasets were analyzed and compared.Next,the existing knowledge selection methods were classified,and the strengths and weaknesses of each category were analyzed.Following that,the existing reasoning methods were categorized and discussed.Finally,the possible future directions in this field were explored.