Cross-lingual Linking of Multi-word Entities and their corresponding Acronyms

Cross-lingual Linking of Multi-word Entities and their corresponding Acronyms Guillaume Jacquet1, Maud Ehrmann2, Ralf Steinberger1, Jaakko Väyrynen1 1 European Commission Joint Research Centre Ispra, Italy {fname}.{lname}@jrc.ec.europa.eu 2 Ecole Polytechnique Fédérale de Lausanne Digital Humanities Laboratory Lausanne, Switzerland maud.ehrmann@epfl.ch Abstract This paper reports on an approach and experiments to automatically build a cross-lingual multi-word entity resource. Starting from a collection of millions of acronym/expansion pairs for 22 languages where expansion variants were grouped into monolingual clusters, we experiment with several aggregation strategies to link these clusters across languages. Aggregation strategies make use of string similarity distances and translation probabilities and they are based on vector space and graph representations. The accuracy of the approach is evaluated against Wikipedia’s redirection and cross-lingual linking tables. The resulting multi-word entity resource contains 64,000 multi-word entities with unique identifiers and their 600,000 multilingual lexical variants. We intend to make this new resource publicly available. Keywords: Multi-Word Named Entity, Named Entity Cross-lingual linking, acronyms 1. Introduction Named entities (NEs) such as persons, organisations, locations and events are major bearers of information in text as they provide answers to the text represen- tation questions Who did What to Whom, Where and When? For this reason, work on Named Entity Recog- nition and Classification is abundant (e.g. Nadeau and Sekine (2007)) and NEs have been linked to knowledge bases (Rao et al., 2013; McNamee and Dang, 2009). Major challenges are homographic entity names be- longing to different classes or within the same class, as well as the existence of variant spellings within the same or across different languages (Steinberger et al., 2011). The situation gets even more complex for multi-word entity names because such names are usually com- posed of words from the common language (e.g. Eco- nomic Community of West African States abbrevi- ated as ECOWAS). These common words are normally translated when referred to in different languages (e.g. in Portuguese Comunidade Económica dos Estados da África Ocidental, abbreviated as CEDEAO) and au- thors frequently abbreviate or change the multi-word forms, either out of negligence or for space reasons (e.g. in English Economic Community of West Africa). While one could argue that such abbreviated or newly created names are wrong, they appear daily in real documents and an information-seeking individual or system would be interested in retrieving documents in which the intention was to refer to the entity of inter- est. In a multilingual large-scale media monitoring envi- ronment such as EMM (Steinberger et al., 2009; Stein- berger et al., 2015)1, we observe an abundant number of spelling variants for entities, including spelling mis- takes (e.g. ‘United Nattions’ when referring to United Nations), inflections (e.g. Birleşmiş Milletler’in where the inflection suffix ’in is added to the Turkish equiva- lence of ‘United Nations’) and variants in other scripts (e.g. Russian Cyrillic Организация Объединённых На- ций). One daily media monitoring task consists in recognising new names and in determining automat- ically whether they are a new name or whether they might be a spelling variant of a name encountered …

Content truncated to 500 words