基于语言模型验证的词义消歧语料获取.pdfVIP

第22 卷 第6 期 中文信息学报 Vol. 22, No. 6 2008 年 11 月 JOU RNAL OF CH INESE INFORMAT ION PR OCESSIN G Nov., 2008 : 100320077 2008) 0620038205 郭宇航, 车万翔, 刘挺 哈尔滨工业大学 计算机科学与技 学院信息检索研究室, 黑龙江 哈尔滨 150001) : 作为一种稀缺资源, 人工标注语料的匮乏限制了有指导词义消歧系统的大规 应用。有人提出了利用目 标词的单义同义词在生语料中自动获取词义消歧语料的方法, 然而, 在某些上下文当中, 用目标词替换这些单义的 同义词并不合适, 从而带来噪声。为此, 笔者使用语言 型过滤这些噪声, 达到净化训练数据, 提高系统性能的目 的。笔者在 Sen seval2 3 国际评测中文采样词词义消歧数据集上进行了实验, 结果表明经过语言 型过滤的词义消 歧系统性能明显高于未经过滤的系统。 : 计算机应用; 中文信息处理; 词义消歧; 语言 型; 噪声过滤 : T P39 1 : A Wor d Sense Disambiguation Corpus Acquisition by Language Model Validation GUO Yu2hang, CHE Wan2x iang, LIU Tin g Informat ion Retr ieval Lab , School of Computer Science and Technology, H arbin Institut e of Technology, H arbin , H eilongj iang 150001, China) Abstr act: The lack of hand2 crafted tr aining data is a crit ical issue for sup ervised word sense disambiguation WSD) sy stems. The monosemou s lexical r elatives sub stitution of target words have been pr oposed to acquire W SD cor pu s from the Web automatically . H owever, in some cases, the monosemou s lexical r elatives cannot b e substit uted by the target word suit ably and t hen noises will be br ought in . W e propose a language models validat ion method t o filter these noises, which can purify the training data, and improve the per for mance accordingly. Our exp eriments on Sen seva2l 3 Chinese lexical sam ple t ask show that the sy stem based on the tr aining data acquired from the Web with language model validat ion achieves better accuracy than the one without language models validat ion . Key wor ds: computer application ; Chinese information pro