Bootstrapping a Hybrid MT System to a New Language Pair

Bootstrapping a Hybrid MT System to a New Language Pair Jo˜ao Rodrigues, Nuno Rendeiro, Andreia Querido, Sanja ˇStajner, Ant´onio Branco Department of Informatics, Faculty of Sciences University of Lisbon, Portugal {joao.rodrigues, nuno.rendeiro, andreia.querido, stajner.sanja, antonio.branco}@di.fc.ul.pt Abstract The usual concern when opting for a rule-based or a hybrid machine translation (MT) system is how much effort is required to adapt the system to a different language pair or a new domain. In this paper, we describe a way of adapting an existing hybrid MT system to a new language pair, and show that such a system can outperform a standard phrase-based statistical machine translation system with an average of 10 persons/month of work. This is specifically important in the case of domain-specific MT for which there is not enough parallel data for training a statistical machine translation system. Keywords: Hybrid Machine Translation, Tecto MT, Portuguese Synthesis 1. Introduction Phrase-based statistical machine translation (PBSMT) is considered as state-of-the-art MT approach whenever suf- ficiently large parallel (or comparable) datasets for training are available. However, for many language pairs and trans- lation directions (English to Portuguese among them) large training datasets only exists for few domains, such as par- liamentary discussions (Europarl (Koehn, 2005)) or legal documents (JRC-Acquis corpus (Steinberger et al., 2006)). In such cases, it is assumed that a rule-based or a hybrid MT system lead to better results as it can better overcome the data sparsity and generalise over the unseen word forms (especially useful in the case of morphologically rich lan- guages). The main concern is, however, that adaptation of a rule-based or a hybrid MT system (its rules) to a new lan- guage pair or a new domain may require considerable time and effort. In this paper, we focus on a hybrid MT system (TectoMT (Popel and ˇZabokrtsk`y, 2010)) and show that the adap- tation of the existing TectoMT system for English-Czech language pair to a new language pair (English to Por- tuguese translation) can be done with an average of 10 per- sons/month of work time and outperform the standard PB- SMT system on a domain-specific MT task. In the next section (Section 2.), we briefly describe the ar- chitecture of the TectoMT system and its main components. Section 3. describes how we adapted from the English syn- thesis to the required Portuguese synthesis. In Section 4. we present the results of the automatic evaluation of our En- glish to Portuguese TectoMT system and discussion. Sec- tion 5. contains closing remarks and directions for future work. 2. Related Work TectoMT is a modular, machine translation system (Popel and ˇZabokrtsk`y, 2010), based on the Prague tectogrammat- ics theory of Sgall (Sgall et al., 1986). It uses two layers of structural description, the shallow a-layer (syntactic analy- sis) and the deep t-layer (Figure 1). The tree-to-tree transfer (ˇZabokrtsk´y et al., 2008) is per- formed on the deep t-layer (tectogramatical layer) by using the Maximum Entropy context-sensitive translation models (Mareˇcek et al., 2010). The analysis phase (which converts the input sentence …

Content truncated to 500 words