TaPaCo

Tatoeba是一个主要针对语言学习者的众包项目。其目的是为特定的语言结构和单词提供例句和翻译。释义语料库是通过在图形中填充Tatoeba句子和句子之间的等效链接 “意思是同一件事” 来创建的。然后遍历该图以提取释义集。应用了几个与语言无关的过滤器和修剪步骤来删除无趣的句子。对三种语言进行的手动评估表明，推断出的释义中有一半到四分之三是正确的，而其余大多数是正确但微不足道的，或者近似于消除形态差异的释义。该语料库总共包含190万个句子，每个语言200-250 000个句子。它涵盖了一系列语言，据我们所知，没有其他解释数据集。

Tatoeba is a crowdsourcing project primarily targeting language learners. Its core goal is to provide example sentences and translations for specific linguistic structures and lexical items. This paraphrase corpus is constructed by building a graph populated with Tatoeba sentences and their equivalence links—i.e., sentence pairs that express the same meaning—then traversing the graph to extract paraphrase sets. Several language-agnostic filtering and pruning procedures are applied to remove uninformative or uninteresting sentences. Manual evaluations conducted on three languages demonstrate that half to three-quarters of the inferred paraphrases are correct, while the majority of the remaining instances are either correct but trivial, or paraphrases that approximate the elimination of morphological discrepancies. In total, the corpus comprises 1.9 million sentences, with 200,000 to 250,000 sentences per individual language. It covers a broad spectrum of languages, and to the best of our knowledge, no other comparable paraphrase datasets are currently available.

搜集汇总

背景与挑战

背景概述

TaPaCo是一个基于Tatoeba众包项目构建的释义语料库，通过图形方法提取句子间的释义关系并经过过滤处理。该数据集覆盖73种语言，包含约190万个句子，每个语言有20-25万句子，手动评估表明其释义准确性较高。