QAMR (Question-Answer Meaning Representation Dataset)

问答意义表示是表示谓词-论元结构的一种新范式，它利用自由形式的问题及其答案来表示广泛的语义现象。 QAMR 的语义表达能力与现有的形式主义相比（在某些情况下超过），而表示可以由非专家注释（特别是使用众包）。 我们定义了 QAMR，开发了一个众包管道，用于在 Mechanical Turk 上大规模收集它们，收集大约 5,000 个带注释的句子的数据集，对这些数据进行彻底分析，并在数据集上运行一些内在基线。除此之外，监督开放信息提取（代码）的并行工作利用 QAMR 数据集来提高开放 IE 系统的性能，尤其是在不太常见的注释谓词（如名词化）上。Question-Answer Meaning Representation (QAMR) is a novel paradigm for representing predicate-argument structure, which leverages free-form questions and their corresponding answers to model a wide range of semantic phenomena. The semantic expressiveness of QAMR is comparable to, and in some cases exceeds, that of existing formalisms, and its annotations can be completed by non-experts, especially via crowdsourcing. We formally define QAMR, develop a crowdsourcing pipeline for large-scale data collection on Amazon Mechanical Turk, construct a dataset containing approximately 5,000 annotated sentences, conduct thorough analysis of the collected data, and run several intrinsic baselines on the dataset. In addition, a parallel work on supervised open information extraction (code) leverages the QAMR dataset to improve the performance of open IE systems, particularly on less frequently annotated predicates such as nominalizations.

搜集汇总

背景与挑战

背景概述

QAMR是一种问答意义表示数据集，采用自由形式问题与答案作为新范式来表征语义现象，通过众包方式收集了约5,000个带注释句子。该数据集被用于提升开放信息提取系统的性能，特别是在处理不常见谓词时。