Design Challenges in Named Entity Transliteration

Proceedings of the 27th International Conference on Computational Linguistics, pages 630–640 Santa Fe, New Mexico, USA, August 20-26, 2018. 630 Design Challenges in Named Entity Transliteration Yuval Merhav∗ Amazon Alexa AI Cambridge, MA merhavy@amazon.com Stephen Ash∗ Amazon AWS AI Seattle, WA ashstep@amazon.com Abstract We analyze some of the fundamental design challenges that impact the development of a mul- tilingual state-of-the-art named entity transliteration system, including curating bilingual named entity datasets and evaluation of multiple transliteration methods. We empirically evaluate the transliteration task using the traditional weighted finite state transducer (WFST) approach against two neural approaches: the encoder-decoder recurrent neural network method and the recent, non-sequential Transformer method. In order to improve availability of bilingual named entity transliteration datasets, we release personal name bilingual dictionaries mined from Wikidata for English to Russian, Hebrew, Arabic, and Japanese Katakana. Our code and dictionaries are publicly available1. 1 Introduction Named entity transliteration is the process of converting a named entity from one language script to another, using the correct characters that represent the entity in the target language. It is an important component in many search and language understanding tasks, such as robust cross-language information retrieval (CLIR) and machine translation (MT), among others. A possible simple transliteration approach is mapping every character (or sequence of characters) in the source language to its most common counterpart in the target language. However, spelling and pronunciation of many languages (e.g., English, Japanese) can be ambiguous and inconsistent (Fushimi et al., 1999). As a result, most transliteration systems are data driven and use context for disambiguation; e.g., (Yan and Nivre, 2016; Haizhou et al., 2004; Ekbal et al., 2006). While transliteration has been a long studied problem, some important aspects received little attention. There is not clear guidance that addresses a number of common design considerations faced when build- ing a robust multilingual transliteration system, such as data representation and the huge gap in results depending on the language pairs and transliteration direction (Rosca and Breuel, 2016). Like many other NLP fields recently, neural transliteration systems have gained popularity. However, it is still unclear if neural systems consistently outperform traditional approaches and what architecture is ideal for this task. For example, in (Yan and Nivre, 2016) the authors applied a stack of convolutional layers and simple recurrent layer on top for English to Chinese transliteration, which achieved competitive results but still below a phrase-based statistical machine translation system. Other works show that bidirectional long short-term memory (LSTM) neural networks and the encoder-decoder architecture (Sutskever et al., 2014) achieve comparable results with WFST based n-gram models that are considered state-of-the- art (Rao et al., 2015; Yao and Zweig, 2015). However, these works only studied grapheme-to-phoneme (G2P) conversion from English to standard English pronunciation sets, such as the CMU pronunciation dictionary (Weide, 2014). Also, there is not empirical evidence comparing the neural encoder-decoder method with LSTM to the recently proposed Transformer (Vaswani et al., 2017) neural method on the This work is licensed under a Creative Commons Attribution 4.0 International …

Content truncated to 500 words