Rich bitext projection features for parse reranking丰富的文本解析投影特征重排序.pptVIP

下载本文档

3
0
约1.99万字
约 45页
2017-03-08 发布于上海
举报
版权申诉

Rich bitext projection features for parse reranking丰富的文本解析投影特征重排序.ppt

1、原创力文档（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。。
2、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击这里二次下载。
3、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。
4、该文档为VIP文档，如果想要下载，成为VIP会员后，下载免费。
5、成为VIP后，下载本文档将扣除1次下载权益。下载后，不支持退款、换文档。如有疑问请联系我们。
6、成为VIP后，您将拥有八大权益，权益包括：VIP文档下载权益、阅读免打扰、文档格式转换、高级专利检索、专属身份标志、高级客服、多端互通、版权登记。
7、VIP文档为合作方或网友上传，每下载1次，网站将根据用户上传文档的质量评分、类型等，对文档贡献者给予高额补贴、流量扶持。如果你也想贡献VIP文档。上传文档

Rich bitext projection features for parse reranking丰富的文本解析投影特征重排序

Dealing with agglutination As I mentioned previously, one problem with Toutanova‘s work is treating agglutination as if it is inflection It is intuitive to instead segment to deal with agglutination We are currently doing this for a common portmanteau in German: Preposition + Article E.g., ?zum“ - this is the preposition ?zu“ and the definite article ?dem“ This means we have to work with a segmented representation (e.g., zu+Dative, definite_article in the stemmed markup) for training and inflection prediction Then synthesize: creation of portmanteaus dependis on the inflection decision Recently, we got this to work for German compounds as well We translate to compound head words and compound non-head words, then subsequently combine them. Finally we inflect them. Evaluation WMT 2009 English to German news task All parallel training data (about 1.5 M parallel sentences, mostly Europarl) Standard Dev and Test sets Two limitations of the experiments here: We were not able to parse the monolingual data, so we are not using it (except in one experiment...) The inflection prediction system that predicts grammatical features does not currently have access to an inflected word form LM We have recently overcome these, see our EACL 2012 paper System BLEU (end-to-end, case sensitive) Baseline 12.62 1 LM predicting surface forms, no portmanteau handling 12.31 1 LM predicting surface forms (11 M sentences inflection prediction training), no portmanteau handling 12.72 1 LM predicting surface forms 12.80 1 LM predicting grammatical features 13.29 4 LMs, each predicting one grammatical feature 13.19 4 CRFs, German features only 13.39 4 CRFs, German and English features 13.58 Newest developments We now have a rule-based preprocessing setup for English to German translation See our EACL 2012 paper This does reordering of English clauses by analyzing what the translated German clause type will be We are currently working on combining inflection, compounding, verbal reordering and