Cross-lingual Named Entity List Search via Transliteration

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4247–4255 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4247 Cross-lingual Named Entity List Search via Transliteration Aleksandr Khakhmovich1,2, Svetlana Pavlova1, Kira Kirillova1, Nikolay Arefyev1,2,3, Ekaterina Savilova1 1Samsung R&D Institute Russia 2Lomonosov Moscow State University 3National Research University Higher School of Economics Moscow, Russian Federation {a.hahmovich, s.pavlova, k.kirillova, n.arefyev, e.savilova}@partner.samsung.com Abstract Out-of-vocabulary words are still a challenge in cross-lingual Natural Language Processing tasks, for which transliteration from source to target language or script is one of the solutions. In this study, we collect a personal name dataset in 445 Wikidata languages (37 scripts), train Transformer-based multilingual transliteration models on 6 high- and 4 less-resourced languages, compare them with bilingual models from (Merhav and Ash, 2018) and determine that multilingual models perform better for less-resourced languages. We discover that intrinsic evaluation, i.e comparison to a single gold standard, might not be appropriate in the task of transliteration due to its high variability. For this reason, we propose using extrinsic evaluation of transliteration via the cross-lingual named entity list search task (e.g. personal name search in contacts list). Our code and datasets are publicly available online. Keywords: Natural Language Processing, Transliteration, Named Entity Transliteration, Cross-Lingual, Multilingual, Transformer 1. Introduction Transliteration, being a common way to adapt the out-of- vocabulary words, such as named entities of any types, from one language (script) to another, is an important component in many language processing tasks. A popular way to ap- proach transliteration is to consider it a specific case of ma- chine translation. Transliteration faces ambiguity problem due to large amount of language-specific phonetic varia- tions. Another challenge, mostly specific for rare or less- resourced languages, lies in the field of data collection. In this study1, we apply a Transformer-based solution for multilingual transliteration as a core technology, and eval- uate it in a cross-lingual named entity list search task. The contributions in this paper are summarized as follows: • We collect and release a multilingual personal name dataset obtained from Wikidata, which covers 445 lan- guages and 37 scripts. Scripts for basic pre-processing and train, dev, test split are supplied. • We propose multilingual transliteration models which outperform bilingual models presented in (Mer- hav and Ash, 2018)2 in English→target lan- guage transliteration experiment on multiple lan- guage pairs: English→Hebrew, English→Japanese, English→Katakana3, English→Korean. We also test the models on 4 new less-resourced languages (Belaru- sian, Odia, Punjabi, Telugu) and obtain better results with multilingual models than with bilingual. • We provide a linguistic analysis of the transliteration model predictions and reveal significant task-specific problems related to transliteration variability. 1Current article repository: https://github.com/perspective- Alex/NE-transliteration-and-search. 2(Merhav and Ash, 2018) article repository: https://github.com/steveash/NETransliteration-COLING2018/ 3Detailed information on the choice to separate out personal names Katakana is provided in Section 3.2 • We propose the cross-lingual named entity list search task, prepare use case-specific evaluation datasets, and use the devised transliteration model to solve it. The rest of the paper is organized …

Content truncated to 500 words