Training Automatic Transliteration Models on DBPedia Data

Proceedings of Recent Advances in Natural Language Processing, pages 654–662, Hissar, Bulgaria, Sep 7–9 2015. Training Automatic Transliteration Models on DBpedia Data Velislava Todorova Linguistic Modeling Department IICT-BAS slava@bultreebank.org Kiril Simov Linguistic Modeling Department IICT-BAS kivs@bultreebank.org Abstract Our goal is to facilitate named en- tity recognition in Bulgarian texts by extending the coverage of DBpedia (http://www.dbpedia.org/) for Bul- garian. For this task we have trained translation Moses models to transliter- ate foreign names to Bulgarian. The training sets were obtained by extract- ing the names of all people, places and organizations from DBpedia and its extension Airpedia (http://www. airpedia.org/). Our approach is ex- tendable to other languages with small DBpedia coverage. 1 Introduction DBpedia Linked Open Dataset1 provides the wealth of Wikipedia in a formalized way via the ontological language in which DBpedia statements are represented. Still DBpedia re- flects the multilingual nature of WikiPedia. But if a user needs to access the huge num- ber of instances extracted from the English version of WikiPedia, for example, in a dif- ferent language (in our case Bulgarian) he/she will not be able to do so, because the DBpedia in the other language will not provide appro- priate Uniform Resource Identifiers (URIs) for many of the instances in the English DBpe- dia. In this paper we describe an approach to the problem. It generates appropriate names in Bulgarian from DBpedia URIs in other lan- guages. These new names are used in two ways: (1) to form gazetteers for annotation of DBpedia instances in Bulgarian texts; and (2) they refer back to the DBpedia URIs from which they have been created and in this way 1http://wiki.dbpedia.org/ provide access to all RDF statements about the original URIs. The paper presents several transliteration models and their evaluation. Their evaluation is done over 100 examples, transliterated man- ually by two people, independently from each other. The discrepancies between the two hu- man transliterations demonstrate the complex- ity of the task. The structure of the paper is as follows: in section 2 we describe the problem in more de- tail; in section 3 we present some related ap- proaches; section 4 reports on the preliminary experiments; section 5 describes how the train- ing data is extended on the basis of the results from the preliminary models and new models are trained; section 6 provides some heuristic rules for combining transliteration and trans- lation of names’ parts; section 7 describes the evaluation of the models; the last section con- cludes the paper and presents some directions for future work. 2 Challenges The transliteration of proper names presents many and quite challenging difficulties not only to automatic systems, but also to humans. A lot of information is needed to perform the task: the language of origin to determine the pronunciation; some real world facts like how this name was/is actually pronounced by the person it belongs or belonged to (in case of personal names) or by the locals (in case of toponyms) and so on; and also the tradition …

Content truncated to 500 words