KILT(KILT Benchmark)

KILT (Knowledge Intensive Language Tasks) 是一个由 11 个数据集组成的基准，代表 5 种类型的任务：事实检查 (FEVER)、实体链接 (AIDA CoNLL-YAGO、WNED-WIKI、WNED-CWEB)、插槽填充 (T-Rex , Zero Shot RE), 开放域 QA (Natural Questions, HotpotQA, TriviaQA, ELI5), Dialog generation (Wizard of Wikipedia)。所有这些数据集都基于一个预处理的维基百科快照，允许更公平和更一致的评估，并支持新的任务设置，例如多任务和迁移学习。

KILT (Knowledge Intensive Language Tasks) is a benchmark composed of 11 datasets covering 5 types of tasks: Fact Checking (FEVER), Entity Linking (AIDA CoNLL-YAGO, WNED-WIKI, WNED-CWEB), Slot Filling (T-Rex, Zero-Shot RE), Open-Domain QA (Natural Questions, HotpotQA, TriviaQA, ELI5), and Dialogue Generation (Wizard of Wikipedia). All these datasets are built on a preprocessed Wikipedia snapshot, enabling fairer and more consistent evaluation, and supporting novel task settings such as multi-task learning and transfer learning.

搜集汇总

背景与挑战

背景概述

KILT是一个知识密集型语言任务基准，包含11个数据集，覆盖5种任务类型，基于预处理的维基百科快照，支持多任务和迁移学习。