Entity Projection via Machine Translation for Cross-Lingual NER

Entity Projection via Machine-Translation for Cross-Lingual NER Alankar Jain Bhargavi Paranjape Carnegie Mellon UniversityZachary C. Lipton Pittsburgh, USA {alankarjain91,bhargavi22294}@gmail.com, zlipton@cmu.edu Abstract Although over 100 languages are supported by strong off-the-shelf machine translation sys- tems, only a subset of them possess large annotated corpora for named entity recogni- tion. Motivated by this fact, we leverage machine translation to improve annotation- projection approaches to cross-lingual named entity recognition. We propose a system that improves over prior entity-projection methods by: (a) leveraging machine translation systems twice: first for translating sentences and sub- sequently for translating entities; (b) match- ing entities based on orthographic and pho- netic similarity; and (c) identifying matches based on distributional statistics derived from the dataset. Our approach improves upon cur- rent state-of-the-art methods for cross-lingual named entity recognition on 5 diverse lan- guages by an average of 4.1 points. Further, our method achieves state-of-the-art F1 scores for Armenian, outperforming even a monolin- gual model trained on Armenian source data.1 1 Introduction While machine learning methods for various Nat- ural Language Processing (NLP) tasks have pro- gressed rapidly, the benefits accrue disproportion- ately among languages endowed with large anno- tated corpora. Owing to the dependence of state- of-the-art deep learning approaches on massive amounts of data, creating suitable datasets can be prohibitively expensive. This asymmetry be- tween resource-rich and relatively under-resourced languages has inspired work on cross-lingual ap- proaches that leverage annotated datasets from the former to build strong models for the latter. This paper focuses on cross-lingual approaches to Named Entity Recognition (NER), owing to 1Code for our paper can be found at: https://github. com/alankarj/cross_lingual_ner NER’s importance as a core component in infor- mation retrieval and question answering systems. Specifically, we focus on medium-resource lan- guages. We define these to be languages for which although annotated NER corpora do not exist, off- the-shelf Machine Translation (MT) systems, such as Google Translate2, do. We are motivated by the fact that although there are fewer than 50 lan- guages for which large NER datasets (greater than 200k tokens) with gold annotations are publicly available3, many more languages are supported by good-quality MT (Wu et al., 2016). Google Trans- late alone supports 103 languages4, many of which have either no, or only small, NER datasets. We address the setting where annotated cor- pora exist in the source (resource-rich) language— English in our experiments—but for the target (medium-resource) language, we can only afford to label a small validation set. We tackle this problem by first creating an unlabeled dataset in the target language by translating each sentence in the source dataset to the target language. For MT, we use Google Translate, motivated by its large coverage. Next, we annotate this dataset via entity projec- tion—first aligning every entity in a source sen- tence with its counterpart in the corresponding tar- get sentence (entity alignment) and then projecting the tags from source to target in the aligned entity pairs (tag projection). One consequence of relying on MT as opposed to word-by-word or phrase-by- phrase …

Content truncated to 500 words