WikiHop

WikiHop 是一个多跳问答数据集。 WikiHop 的查询由 WikiData 中的实体和关系构成，而支持文档则来自 WikiReading。首先构建一个连接实体和文档的二分图，每个查询的答案通过遍历该图来定位。包括与答案类型一致并且在查询中与答案共享相同关系的候选者，从而产生一组候选者。因此，WikiHop 是一个多选风格的阅读理解数据集。训练集中大约有 43K 样本，开发集中有 5K 样本，测试集中有 2.5K 样本。未提供测试集。任务是在给定查询和多个支持文档的情况下预测正确答案。该数据集包括一个掩码变体，其中所有候选者及其在支持文档中的提及都被随机但一致的占位符标记替换。

WikiHop is a multi-hop question answering dataset. Its queries are constructed from entities and relations in WikiData, while its supporting documents are sourced from WikiReading. First, a bipartite graph connecting entities and documents is built, and the answer to each query is located by traversing this graph. Candidates are generated by including those that match the answer type and share the same relation with the answer as stated in the query, thus forming a candidate set. Therefore, WikiHop is a multiple-choice-style reading comprehension dataset. There are approximately 43K samples in the training set, 5K in the development set, and 2.5K in the test set; however, the test set is not provided. The task is to predict the correct answer given the query and multiple supporting documents. The dataset also includes a masked variant, where all candidates and their mentions in the supporting documents are replaced with random but consistent placeholder tokens.

搜集汇总

背景与挑战

背景概述

WikiHop是一个多跳问答数据集，基于WikiData实体和关系构建查询，支持文档来自WikiReading，采用多选阅读理解形式。数据集包含训练集、开发集和测试集（未公开），并提供一个掩码变体用于替换候选者及其提及。