Mitigating Problems in Analogy-based EBMT with SMT and vice versa: A Case Study with Named Entity Transliteration

Mitigating Problems in Analogy-based EBMT with SMT and vice versa: a Case Study with Named Entity Transliteration∗ Sandipan Dandapat, Sara Morrissey, Sudip Kumar Naskar, and Harold Somers Centre for Next Generation Localisation Dublin City University Glasnevin, Dublin 9, Ireland {sdandapat, smorri, snaskar, hsomers}@computing.dcu.ie Abstract. Five years ago, a number of papers reported an experimental implementation of an Example Based Machine Translation (EBMT) system using proportional analogy. This approach, a type of analogical learning, was attractive because of its simplicity; and the paper reported considerable success with the method using various language pairs. In this paper, we describe our attempt to use this approach for tackling English–Hindi Named Entity (NE) Transliteration. We have implemented our own EBMT system using proportional analogy and have found that the analogy-based system on its own has low precision but a high recall due to the fact that a large number of names are untransliterated with the approach. However, mitigating problems in analogy-based EBMT with SMT and vice-versa have shown considerable improvement over the individual approach. Keywords: Example Based Machine Translation, Proportional Analogy, Named Entity Transliteration ∗ The research is supported by the Science Foundation Ireland (Grant 07/CE/I1142) as part of the Centre for Next Generation Localization (http://www.cngl.ie) at Dublin City University. Copyright 2010 by Sandipan Dandapat, Sara Morrissey, Sudip Kumar Naskar, and Harold Somers 1 Introduction In the EBMT workshop in Phuket, Thailand, Lepage and Denoual (2005a) presented “The ‘purest’ EBMT system ever built: no variable, no templates, no training, examples, just examples, only examples”. This purely data-driven approach to MT uses the notion of proportional analogy (PA, described below), a type of analogical learning, the very simplicity of which is its attraction. We have attempted to reimplement the method in order to verify the previous authors’ claims. It has been noted, that the PA-based system works well in the case of shorter sentences with similar structure (Langlais and Patry, 2007). Thus we have taken named entity (NE) transliteration as a case study because NEs are typically short. Furthermore, it was reported (Hermjakob et. al, 2008) that the state-of-the-art SMT system can’t handle NEs that are not found in the training parallel text. In this paper we describe an experiment in which we use the method to handle unknown named entities (NEs) in an English-Hindi transliteration task. The idea introduced in Lepage and Denoual (2005a) is explained in considerably more detail in Lepage and Denoual (2005b, c). ALEPH system is an implementation of the research described in their three 2005 papers, and was tested on a corpus of 160K English, Japanese and Chinese sentences, from the C-STAR project’s BTEC, a travel and tourism domain corpus. The system did very well on data from the IWSLT 2004 competition, coming a close second to the competition winner on all measures. The ALEPH system evolved into a new system, named GREYC, with some modification described in Lepage and Lardilleux (2007). The GREYC system also incorporated new heuristics and had an additional refinement of non-determinism to generate all possible …

Content truncated to 500 words