A Novel Method to Evaluate Romanization Systems: The Case of Romanizing Arabic Proper Nouns

The transliteration of Arabic proper nouns to other languages is usually based on the phonetic translation of these nouns into their phonetic Latin counterparts. Most of the dictionaries do not include most of these nouns, although some may have meanings. Transliteration is essential generally to Natural Language Processing (NLP) field and specifically to machine translation systems, cross-language information-retrieval systems and Web search engines, since most of the submitted queries are proper nouns. Romanization also known as Latinization that refers to the representation of names and technical terms with Roman (Latin) alphabets. Romanization is accomplished based on a number of methods. Machine Romanizers are either based on a dictionary or on rules to Romanize different proper nouns. The Romanization process is not trivial due to many problems, so normally there are different Roman variations of a single Arabic proper noun that cause many problems to search engines and databases. This study is based on a dataset consisting of around 5000 Arabic proper nouns as a case study, where this dataset accumulated through years as a result of indexing the authors of Arabic books at the library of one of the public universities in the Middle East. Many methods are presented to evaluate Romanization systems. In this paper, we presented a new automated evaluation method called, "Back Romanization" to automatically evaluate the effectiveness of different Romanization systems regardless of the source and destination languages used.