Traditional statistical machine translation (SMT) architectures, like the one implemented in this chapter, address the translation task as a search problem (Brown et al., 1990). Given an input string in the source language, the goal is to find the output string in the target language which maximizes the product of a series of probability models over the search space defined by all possible partitions of the source string and all possible reorderings of the translated units. This search process implicitly decomposes the translation problem into two separate but interrelated subproblems:

Word selection, also referred to as lexical choice, is the problem of deciding, given a word (or phrase) f in the source sentence, which word (or phrase) e in the target language is the most appropriate translation. This problem is mainly addressed by translation models, which serve as probabilistic bilingual dictionaries, typically accounting for P(f|e), P(e|f) of P(e, f). Translation models provide, for each word (or phrase) in the source vocabulary, a list of translation candidates with associated translation probabilities. During the search there is another component which addresses word selection, the language model. This component helps the decoder to move toward translations which are more appropriate, in terms of grammaticality, in the context of what is known so far about the target sentence being generated.

Word ordering refers to the problem of deciding which position the translation candidate e must occupy in the target sentence. This problem is mainly addressed by the reordering model which allows for certain word movement inside the sentence. Again, the language model helps the decoder, in this case, to move toward translations which preserve a better word ordering according to the rules of the target language.

In standard phrase-based SMT systems, like that described by Koehn et al. (2003), the estimation of these models is fairly simple. For instance, translation models are built on the basis of relative frequency counts, i.e., maximum likelihood estimates (MLEs). Thus, all the occurrences of the same source phrase are assigned, no matter what the context is, the same set of translation probabilities. For that reason, recently, there is a growing interest in the application of discriminative learning, both for word ordering (Chang and Toutanova, 2007; Cowan et al., 2006) and, especially, for word selection (Bangalore et al., 2007; Carpuat and Wu, 2007b; Giménez and Màrquez, 2007a; Stroppa et al., 2007; Vickrey et al., 2005).

Interest in discriminative word selection has also been motivated by recent results in word sense disambiguation (WSD). The reason is that SMT systems perform an implicit kind of WSD, except that instead of working with word senses, SMT systems operate directly on their potential translations. Indeed, recent semantic evaluation campaigns have treated word selection as a separate task, under the name of multilingual lexical sample (Chklovski et al., 2004; Jin et al., 2007). Therefore, the same discriminative approaches that have been successfully applied to WSD should be also applicable to SMT. In that spirit, instead of relying on MLE for the construction of the translation models, approaches to discriminative word selection suggest building dedicated discriminative translation models which are able to take a wider feature context into account. Lexical selection is, therefore, addressed as a classification task. For each possible source word (or phrase) according to a given

bilingual lexical inventory (e.g., the translation model), a distinct classifier is trained to predict lexical correspondences based on local context. Thus, during decoding, for every distinct instance of every source phrase, a distinct context-aware translation probability distribution is potentially available.

Here we extend the work presented in Giménez and Màrquez (2007a). First, in section 11.2, we describe previous and current approaches to dedicated word selection. Then, in section 11.3, our approach to discriminative phrase translation (DPT) is fully described. We present experimental results on the application of DPT models to the Spanish-to-English translation of European Parliament proceedings. In section 11.4, prior to considering the full translation task, we measure the local accuracy of DPT classifiers at the isolated phrase translation task in which the goal is not to translate the whole sentence but only individual phrases without having to integrate their translations in the context of the target sentence. We present a comparative study on the performance of four different classification settings based on two different learning paradigms, namely support vector machines and maximum entropy models.

We shall tackle the full translation task. We have built a state-of-theart factored phrase-based SMT system based on linguistic data views at the level of shallow parsing (Giménez and Màrquez, 2005, 2006). We compare the performance of DPT- and MLE-based translation models built on the same parallel corpus and phrase alignments. DPT predictions are integrated into the SMT system in a soft manner, by making them available to the decoder as an additional log-linear feature so they can fully interact with other models (e.g., language, distortion, word penalty, and additional translation models) during the search. We separately study the effects of using DPT predictions for all phrases as compared to focusing on a small set of very frequent phrases.

This chapter has also served to study the problem of machine translation evaluation. We have applied a novel methodology for heterogeneous automatic MT evaluation which allows for separately analyzing quality aspects at different linguistic levels, e.g., lexical, syntactic, and semantic (Giménez and Màrquez, 2007b). This methodology also offers a robust mechanism to combine different similarity metrics into a single measure of quality based on human likeness (Giménez and Màrquez, 2008). We have complemented automatic evaluation results through error analysis and by conducting a number of manual evaluations.