Many early neural Information Retrieval (NeurIR) methods are re-rankers that rely on a traditional first-stage retriever due to expensive query time computations. Recently, representation-based retrievers have gained much attention, which learns query representation and document representation separately, making it possible to pre-compute document representations offline and reduce the workload at query time. Both dense and sparse representation-based retrievers have been explored. However, these methods focus on finding the representation that best represents a text (aka metric learning) and the actual retrieval function that is responsible for similarity matching between query and document is kept at a minimum by using dot product. One drawback is that unlike traditional term-level inverted index, the index formed by these embeddings cannot be easily re-used by another retrieval method. Another drawback is that keeping the interaction at minimum hurts retrieval effectiveness. On the contrary, interaction-based retrievers are known for their better retrieval effectiveness. In this paper, we propose a novel SEgment-based Neural Indexing method, SEINE, which provides a general indexing framework that can flexibly support a variety of interaction-based neural retrieval methods. We emphasize on a careful decomposition of common components in existing neural retrieval methods and propose to use segment-level inverted index to store the atomic query-document interaction values. Experiments on LETOR MQ2007 and MQ2008 datasets show that our indexing method can accelerate multiple neural retrieval methods up to 28-times faster without sacrificing much effectiveness.

翻译：许多早期的神经信息检索方法由于查询时计算开销大，依赖于传统的第一阶段检索器作为重排序器。近年来，基于表征的检索器受到广泛关注，这类方法分别学习查询表征和文档表征，从而可以离线预计算文档表征并减少查询时的计算负载。基于密集表征和稀疏表征的检索器都已得到探索。然而，这些方法侧重于寻找最能代表文本的表征（即度量学习），而实际负责查询与文档间相似度匹配的检索函数仅通过点积运算实现最小化处理。其弊端在于：与传统词级倒排索引不同，这些嵌入形成的索引难以被其他检索方法复用；同时，最小化交互过程会损害检索效果。相比之下，基于交互的检索器以更优的检索效果著称。本文提出一种新颖的基于片段的神经索引方法SEINE，它构建了通用索引框架，能灵活支持多种基于交互的神经检索方法。我们着重对现有神经检索方法中的常见组件进行细致分解，并提出采用片段级倒排索引存储查询-文档原子交互值。在LETOR MQ2007和MQ2008数据集上的实验表明，本索引方法可在不显著牺牲检索效果的前提下，将多种神经检索方法加速最高达28倍。