2017. cheap translation for Cross-lingual NER 阅读笔记

本文主要是介绍2017. cheap translation for Cross-lingual NER 阅读笔记，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

cheap translation for Cross-lingual NER, Illinois Champaign

• 提出了一个生成翻译字典的 cheap translation 算法
• 该算法可以和 wikifier features、Brown Cluster features 等结合取得更好的效果
• 通过实验说明当 source Lan. 与 traget Lan. 相似度比较高时，可以进一步提高模型性能。
• 探究了 cheap translation 所需字典的size，及其他一些 lan. specific 的 techniques 对模型性能的影响

Abstract

• 目标
  • 面向 low-resource Lan. 的 NER
  • 仅依赖 very minimal resources.
• 方法概述
  • 利用 lexicon translate: annotated high-resource Lan. => target Lan. (training data)
  • 再利用 translated target data 学习一个 monolingual NER model
• 结果
  • 当 target lan. 的 wilipedia 数据可获得时，本文的方法可以提升 wikipedia-based methods 得到 SOA 的结果。

Cheap Translation

• core idea: translate annotated high-re. Lan. => target Lan.
  • using: a lexicon, rather than parallel text
  • challenge: a source word <=> a phrase translation table in target Lan.
    • prominence score: normalized coocurrence counts in each candidate set
    • greeding decoding method
    • choose a candidate, copy labels to it
    • options from the lexicon are resolved by a LM multiplied by the prominence score of each option (Stolcke, 2002)
  • Caution: no reordering of the translated result
• Result
  • annotated data in target Lan.
  • in spite of mistakes, the context around the entities (matters for NER) is reasonably well-preserved

NER Model

• Illinoise NER system
• standard features: forms, capitalization, affixes, word prior, word after, etc.
• Brown cluster features
• multilingual gazetteers
• wiki features

Experiment

baseline

• tranin on English => apply the model directly to the target Lan.
• no gazetteers
• result
  • dosen’t work on non-Latin lan.

cheap translation

• translate English to target lan. (lexicon coverage)
• Yoruba
  • normalize the text (remove all pronunciation sys.) => make the data less sparse
• Bengali & Tamil
  • omit using the word suface as a feature
• result
  • 14.6 points of F1 improv. over the baseline
  • proposed approach is othogonal to other approaches, and can be combined with great effects.

+ Wikifier Features

• obtained by grounding words & phrases to English Wikipedia pages
• using the categories of the linked page as NER features for the surface text.
• result
  • improve scores for all 7 lan.
  • improvement at European Lan > other Lan. => it’s advantageous to select a source lan. similar to the target lan.

Google Translate

• translate English into other Lan. by Google
• aligne the source-target data using fast align
  • the alignments can be noisy given the relatively small amount of text
• project labels across alignments
• result
  • “As with the other approaches, Brown cluster features are an important signal.”

Translation from Similar Lan.

• required 2 new resources
  • annotated data in a similar Lan. S
    • “it is likely that there exists an annotated dataset in a closer language.”
  • a lexicon: S -> T (target Lan.)
    • collect all English words that appear in both dictionaries => two sets of candidate translations
      • too many entries, some incorrect, but correct entry is there.
• choice of the source Lan: WALS
• result
  • 5.4 points higher than Enlish-source

Translation from Similar Lan. + Wikifeier features

• result
  • best cross-lan. setting scores.

Other Expt.

Dictionary Ablation: what effect the size of the lexicon has on the end result?

• only a small numbe rof dictionary entries are likely to be useful
• a small but carefully constructed manual dictionary could have a large impact.

Uyghur

• no wikifier features: Wikipedia size is too small
• Dictionary: LORELEI + Rolston => 116k
• Name Substitution: untranslatable => replace with a randomly selected NE from the gazetteer list corresponding to the tag. => sentence is not fluent, but NEs are fluent in the target text.
• Stemming (词干分析): remove all possible suffixes
• omit form feature for Bengali & Tamil
• result
  • stemming helps => makes the features more dense

+ generate dicionaries using observations over Yyghur and Uzbek

• manual mapping: 100 words
• edit-distance mapping: a modified edit-distance algo.
• cross-lingual CCA with word vectors
  • using CCA project Uyghur and Uzbed monolingual vctor into a share semantic space
  • closest Uyghur word to each Uzbek word.
• result
  • lang. spec. no combine: not much
  • combine ALL lang. specific techs: 10 points higher
  • different training data covers many angles

这篇关于2017. cheap translation for Cross-lingual NER 阅读笔记的文章就介绍到这儿，希望我们推荐的文章对编程师们有所帮助！

---