[Paper Review] "Translation can't change a name": Using Multilingual Data for Named Entity Recognition

Manaal Faruqui|arXiv (Cornell University)|May 4, 2014

Topic Modeling21 references3 citations

TL;DR

This paper proposes using unsupervised word clusters from multilingual, monolingual corpora in the same script to improve named entity recognition (NER) in low-resource settings. By leveraging orthographic stability of named entities—especially person and location names—across related languages, the method significantly boosts NER performance, with the greatest gains observed for phylogenetically close languages and in-domain data.

ABSTRACT

Named Entities (NEs) are often written with no orthographic changes across different languages that share a common alphabet. We show that this can be leveraged so as to improve named entity recognition (NER) by using unsupervised word clusters from secondary languages as features in state-of-the-art discriminative NER systems. We observe significant increases in performance, finding that person and location identification is particularly improved, and that phylogenetically close languages provide more valuable features than more distant languages.

Motivation & Objective

• Address the scarcity of manually annotated training data for named entity recognition (NER) in low-resource languages.
• Investigate whether unsupervised word clusters from secondary languages can improve NER performance in the target language.
• Explore the impact of linguistic relatedness and data domain on cross-lingual transfer effectiveness for NER.
• Demonstrate that orthographically stable named entities (e.g., person and location names) can serve as effective cross-lingual features.
• Evaluate the method across multiple languages (English, German, Dutch, Spanish) using standard NER benchmarks.

Proposed method

• Train unsupervised word clusters on large monolingual corpora from secondary languages using distributional and morphological similarity features.
• Use the resulting word clusters as additional features in a discriminative NER system, specifically a linear-chain Conditional Random Field (CRF).
• Incorporate word clusters from languages sharing the same script (e.g., German, English, French) to improve generalization to out-of-vocabulary named entities.
• Leverage the fact that proper nouns like 'Barack Obama' retain their spelling across orthographically similar languages to transfer knowledge.
• Apply the method using both in-domain (CoNLL) and out-of-domain (WMT-2012 news commentary) monolingual corpora for cluster training.
• Evaluate performance using F1 scores on standard NER test sets, comparing models with and without cross-lingual word cluster features.

Experimental results

Research questions

• RQ1 Can unsupervised word clusters from secondary languages improve NER performance in a target language with limited labeled data?
• RQ2 Does the linguistic relatedness (phylogenetic closeness) between the target and secondary language affect the performance gain from cross-lingual clusters?
• RQ3 How does the domain of the secondary language corpus (in-domain vs. out-of-domain) influence the effectiveness of the transferred features?
• RQ4 Which named entity types (person, location, organization, miscellaneous) benefit most from cross-lingual word clusters?
• RQ5 To what extent do orthographically stable named entities (e.g., proper names) contribute to improved NER generalization across languages?

Key findings

• The use of word clusters from secondary languages significantly improves NER F1 scores, with an average gain of 2.75 F1 points for person entities and 2.25 for location entities.
• The highest performance gains were observed for person (PER) and location (LOC) entities, while miscellaneous (MISC) entities showed minimal improvement due to lexical variation across languages.
• Phylogenetically close languages (e.g., German and Dutch) provided more effective features than distant ones, indicating that linguistic similarity enhances transferability.
• In-domain secondary corpora (e.g., CoNLL data) yielded significantly higher improvements than out-of-domain data (e.g., WMT-2012), highlighting the importance of domain alignment.
• A substantial number of out-of-vocabulary (OOV) named entities in the test set were covered by word clusters trained on secondary languages, particularly for person and location names.
• The method achieved statistically significant improvements across all evaluated languages (English, German, Dutch, Spanish), with the most notable gains in German and Dutch NER systems.