基于多译文的中文转述语料库建设及转述评价方案  

Multi-translation Based Chinese Paraphrase:Evaluation Metric and Corpus

作　　者：阮翀 施文娴 李岩昊 翁伊嘉 胡俊峰[1,2] RUAN Chong;SHI Wenxian;LI Yanhao;WENG Yijia;HU Junfeng(MOE Key Laboratory of Computational Linguistics,Peking University,Beijing 100871,China;School of Electronics and Computer Science,Peking University,Beijing 100871,China)

机构地区：[1]北京大学计算语言学教育部重点实验室,北京100871 [2]北京大学信息科学技术学院,北京100871

出　　处：《中文信息学报》2018年第12期67-75,共9页Journal of Chinese Information Processing

基　　金：国家自然科学基金(61472017)

摘　　要：转述语料是转述现象研究的基础。针对目前学术界中文转述语料稀缺的现状,该文以《简爱》的多个中文译本为基础,通过句对齐得到五万句级别的平行转述语料(1)。使用无监督的小句对齐和词对齐算法,从语料中挖掘到九千多对词汇转述知识。同时,还复现和改进了机器翻译测评指标Meteor,使得该指标更适合于中文转述句子的测评,并构造了一个中文句子转述测评数据集,以便对不同的转述知识和评价指标进行比较。实验表明,该文算法挖掘到的词汇转述知识在封闭测试中不逊于《同义词词林》。Paraphrase corpus is fundamental to research in paraphrase phenomenon,while Chinese paraphrase corpus is hardly available in academia.In this paper,we collected multiple Chinese translations of the novel Jane Eyre,obtaining roughly 50 000 parallel paraphrasing sentences.Then,we managed to extract more than 9 000 pairs of lexical paraphrase knowledge.We further modified METEOR,an automatic machine translation evaluation metric,to better evaluate Chinese paraphrase quality and provided a Chinese paraphrase evaluation dataset.The close test proved a better quality of our mined knowledge than that of Tongyici Cilin.

关 键 词：转述知识挖掘 转述评价指标 转述语料库建设 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]