Named Entity Recognition with Bilingual Constraints

Proceedings of NAACL-HLT 2013, pages 52–62, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Named Entity Recognition with Bilingual Constraints Wanxiang Che† Mengqiu Wang‡ Christopher D. Manning‡ Ting Liu† †{car, tliu}@ir.hit.edu.cn School of Computer Science and Technology Harbin Institute of Technology Harbin, China, 150001 ‡{mengqiu, manning}@stanford.edu Computer Science Department Stanford University Stanford, CA, 94305 Abstract Different languages contain complementary cues about entities, which can be used to im- prove Named Entity Recognition (NER) sys- tems. We propose a method that formu- lates the problem of exploring such signals on unannotated bilingual text as a simple Inte- ger Linear Program, which encourages entity tags to agree via bilingual constraints. Bilin- gual NER experiments on the large OntoNotes 4.0 Chinese-English corpus show that the pro- posed method can improve strong baselines for both Chinese and English. In particular, Chinese performance improves by over 5% absolute F1 score. We can then annotate a large amount of bilingual text (80k sentence pairs) using our method, and add it as up- training data to the original monolingual NER training corpus. The Chinese model retrained on this new combined dataset outperforms the strong baseline by over 3% F1 score. 1 Introduction Named Entity Recognition (NER) is an important task for many applications, such as information ex- traction and machine translation. State-of-the-art su- pervised NER methods require large amounts of an- notated data, which are difficult and expensive to produce manually, especially for resource-poor lan- guages. A promising approach for improving NER per- formance without annotating more data is to exploit unannotated bilingual text (bitext), which are rela- tively easy to obtain for many language pairs, bor- rowing from the resources made available by statis- tical machine translation research.1 Different lan- guages contain complementary cues about entities. For example, in Figure 1, the word “本 (Ben)” is common in Chinese but rarely appears as a trans- lated foreign name. However, its aligned word on the English side (“Ben”) provides a strong clue that this is a person name. Judicious use of this type of bilingual cues can help to recognize errors a mono- lingual tagger would make, allowing us to produce more accurately tagged bitext. Each side of the tagged bitext can then be used to expand the orig- inal monolingual training dataset, which may lead to higher accuracy in the monolingual taggers. Previous work such as Li et al. (2012) and Kim et al. (2012) demonstrated that bilingual corpus an- notated with NER labels can be used to improve monolingual tagger performance. But a major draw- back of their approaches are the need for manual annotation efforts to create such corpora. To avoid this requirement, Burkett et al. (2010) suggested a “multi-view” learning scheme based on re-ranking. Noisy output of a “strong” tagger is used as training data to learn parameters of a log-linear re-ranking model with additional bilingual features, simulated by a “weak” tagger. The learned parameters are then reused with the “strong” tagger to re-rank its own outputs for unseen inputs. Designing …

Content truncated to 500 words