UdS-Sant: English–German Hybrid Machine Translation System

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 152–157, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. UdS-Sant: English–German Hybrid Machine Translation System Santanu Pal1, Sudip Kumar Naskar2, Josef van Genabith1 1Universit¨at des Saarlandes, Saarbr¨ucken, Germany 2 Jadavpur University, Kolkata, India {santanu.pal, josef.vangenabith}@uni-saarland.de sudip.naskar@cse.jdvu.ac.in Abstract This paper describes the UdS-Sant English–German Hybrid Machine Trans- lation (MT) system submitted to the Translation Task organized in the Work- shop on Statistical Machine Translation (WMT) 2015. Our proposed hybrid system brings improvements over the baseline system by incorporating ad- ditional knowledge such as extracted bilingual named entities and bilingual phrase pairs induced from example-based methods. The reported final submission is the result of a hybrid system obtained from confusion network based system combination that combines the best per- formance of each individual system in a multi-engine pipeline. 1 Introduction In this paper, we present Universit¨at des Saarlan- des (UdS) submission (named UdS-Sant) to WMT 2015 using a Hybrid MT framework. We partici- pated in the generic translation shared task for the English-German (EN-DE) language pair. Corpus-based MT (CBMT) has delivered pro- gressively improved quality translations since its inception. There are two main approaches to corpus-based MT – Example Based Machine Translation (EBMT) (Carl and Way, 2003) and Statistical Machine Translation (SMT) (Brown et al., 1993; Koehn, 2010). Out of these two, in terms of large-scale evaluations, SMT is the most suc- cessful MT paradigm. However, each approach has its own advantages and disadvantages along with its own methods of applying and acquiring translation knowledge from the bilingual parallel training data. EBMT phrases tend to be more lin- guistically motivated compared to SMT phrases which essentially operate on n-grams. The knowl- edge extraction as well as representation process, in both EBMT and SMT, uses very different tech- niques in order to extract resources. Even though, SMT is the most popular MT paradigm, it some- times fails to deliver sufficient quality in transla- tion output for some languages, since each lan- guage has its own difficulties. Multiword Expressions (MWEs) and Named Entities (NEs) offer challenges within a language. MWEs are defined as idiosyncratic interpretations that cross word boundaries (Sag et al., 2002). Named entities on the other hand often consist of more than one word, so that they can be consid- ered as a specific type of MWEs such as noun compounds (Jackendoff, 1997). Traditional ap- proaches to word alignment such as IBM Mod- els (Brown et al., 1993) are unable to tackle NEs and MWEs properly due to their inability to han- dle many-to-many alignments. In another well- known word alignment approach, Hidden Markov Model (HMM: (Vogel et al., 1996)), the alignment probabilities depend on the alignment position of the previous word. It does not explicitly consider many-to-many alignment either. We address this alignment problem indirectly. The objective of the present work is threefold. Firstly, we would like to determine how treat- ment of MWEs as a single unit affects the over- all MT quality (Pal et al., 2010; Pal …

Content truncated to 500 words