<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3125">
  <Title>Adri a de Gispert</Title>
  <Section position="3" start_page="0" end_page="162" type="metho">
    <SectionTitle>
2 Baseline N-gram-based SMT System
</SectionTitle>
    <Paragraph position="0"> As already mentioned, the translation model used here is based on bilingual n-grams. It actually constitutes a language model of bilingual units, referred to as tuples, which approximates the joint probability between source and target languages by using bilingual n-grams (de Gispert and Mari no, 2002).</Paragraph>
    <Paragraph position="1"> Tuples are extracted from a word-to-word aligned corpus according to the following two constraints: rst, tuple extraction should produce a monotonic segmentation of bilingual sentence pairs; and second, no smaller tuples can be extracted without violating the previous constraint. See (Crego et al., 2004) for further details.</Paragraph>
    <Paragraph position="2"> For all experiments presented here, the translation model consisted of a 4-gram language model of tuples. In addition to this bilingual n-gram translation model, the baseline system implements a log linear combination of ve feature functions.</Paragraph>
    <Paragraph position="3">  These ve additional models are: A target language model. 5-gram of the target side of the bilingual corpus.</Paragraph>
    <Paragraph position="4"> A word bonus. Based on the number of target words in the partial-translation hypothesis, to compensate the LM preference for short sentences. null A Source-to-target lexicon model. Based on IBM Model 1 lexical parameters(Brown et al., 1993), providing a complementary probability for each tuple in the translation table. These parameters are obtained from source-to-target alignments.</Paragraph>
    <Paragraph position="5"> A Target-to-source lexicon model. Analogous to the previous feature, but obtained from target-to-source alignments.</Paragraph>
    <Paragraph position="6"> A Tagged (POS) target language model. This feature implements a 5-gram language model of target POS-tags. In this case, each translation unit carried the information of its target side POS-tags, though this is not used for translation model estimation (only in order to evaluate the target POS language model at decoding time). Due to the non-availability of POS-taggers for French and German, it was not possible to incorporate this feature in all translation tasks considered, being only used for those translation tasks with Spanish and English as target languages.</Paragraph>
    <Paragraph position="7"> The search engine for this translation system is described in (Crego et al., 2005) and implements a beam-search strategy based on dynamic programming, taking into account all feature functions described above, along with the bilingual n-gram translation model. Monotone search is performed, including histogram and threshold pruning and hypothesis recombination.</Paragraph>
    <Paragraph position="8"> An optimization tool, which is based on a downhill simplex method was developed and used for computing log-linear weights for each of the feature functions. This algorithm adjusts the weights so that a non-linear combination of BLEU and NIST scores is maximized over the development set for each of the six translation directions considered.</Paragraph>
    <Paragraph position="9"> This baseline system is actually very similar to the system used for last year's shared task Exploiting Parallel Texts for Statistical Machine Translation of ACL'05 Workshop on Building and Using Parallel Texts: Data-Driven Machine Translation and Beyond (Banchs et al., 2005), whose results are available at: http://www.statmt.org/wpt05/ mt-shared-task/. A more detailed description of the system can be found in (2005).</Paragraph>
    <Paragraph position="10"> The tools used for POS-tagging were Freeling (Carreras et al., 2004) for Spanish and TnT (Brants, 2000) for English. All language models were estimated using the SRI language modeling toolkit. Word-to-word alignments were extracted with GIZA++. Improvements in word-to-word alignments were achieved through verb group classi cation as described in (de Gispert, 2005).</Paragraph>
  </Section>
  <Section position="4" start_page="162" end_page="163" type="metho">
    <SectionTitle>
3 Reordering Framework
</SectionTitle>
    <Paragraph position="0"> In this section we outline the reordering framework used for the experiments (Crego and Mari no, 2006).</Paragraph>
    <Paragraph position="1"> A highly constrained reordered search is performed by means of a set of reordering patterns (linguistically motivated rewrite patterns) which are used to extend the monotone search graph with additional arcs.</Paragraph>
    <Paragraph position="2"> To extract patterns, we use the word-to-word alignments (the union of both alignment directions) and source-side POS tags. The main procedure consists of identifying all crossings produced in the  is achieved through the POS tags. Three instances of different patterns are extracted using the sentences in the example.</Paragraph>
    <Paragraph position="3">  word-to-word alignments. Once a crossing has been detected, its source POS tags and alignments are used to account for a new instance of pattern. The target side of a pattern (source-side positions after reordering), is computed using the original order of the target words to which the source words are aligned. See gure 1 for a clarifying example of pattern extraction.</Paragraph>
    <Paragraph position="4"> The monotone search graph is extended with re-orderings following the patterns found in training. The procedure identi es rst the sequences of words in the input sentence that match any available pattern. Then, each of the matchings implies the addition of an arc into the search graph (encoding the reordering learnt in the pattern). However, this addition of a new arc is not performed if a translation unit with the same source-side words already exists in the training. Figure 2 shows an example of the  to the original monotone graph (bold arcs) given the reordering patterns found matching any of the source POS tags sequence.</Paragraph>
    <Paragraph position="5"> Once the search graph is built, the decoder traverses the graph looking for the best translation. Hence, the winner hypothesis is computed using all the available information (the whole SMT models). The reordering strategy is additionally supported by a 5-gram language model of reordered source POS-tags. In training, POS-tags are re-ordered according with the extracted reordering patterns and word-to-word links. The resulting sequence of source POS-tags are used to train the n-gram LM.</Paragraph>
    <Paragraph position="6"> Notice that this reordering framework has only been used for some translation tasks (Spanishto-English, English-to-Spanish and English-to-French). The reason is double: rst, because we did not have available a French POS-tagger. Second, because the technique used to learn reorderings (detailed below) does not seem to apply for language pairs like German-English, because the agglutinative characteristic of German (words are formed by joining morphemes together).</Paragraph>
    <Paragraph position="7">  Table 1 shows the improvement of the original baseline system described in section 2 (base), enhanced using reordering graphs (+rgraph) and provided the tagged-source language model (+pos). The experiments in table 1 were not carried out over the of cial corpus of this shared task. The Spanish-English corpus of the TC-Star 2005 Evaluation was used. Due to the high similarities between both corpus (this shared task corpus consists of a subset of the whole corpus used in the TC-Star 2005 Evaluation), it makes sense to think that comparable results would be obtained.</Paragraph>
    <Paragraph position="8"> It is worth mentioning that the of cial corpus of the shared task (HLT-NAACL 2006) was used when building and tuning the present shared task system.</Paragraph>
  </Section>
  <Section position="5" start_page="163" end_page="164" type="metho">
    <SectionTitle>
4 Shared Task Results
</SectionTitle>
    <Paragraph position="0"> The data provided for this shared task corresponds to a subset of the of cial transcriptions of the European Parliament Plenary Sessions. The development set used to tune the system consists of a subset (500 rst sentences) of the of cial development set made available for the Shared Task.</Paragraph>
    <Paragraph position="1">  Table 2 presents the BLEU, NIST and mWER scores obtained for the development-test data set. The last column shows whether the target POS language model feature was used or not. Computed scores are case sensitive and compare to one reference translation. Tasks in bold were conducted allowing for the reordering framework. For French-to-English task, block reordering strategy was used, which is described in (Costa-juss a et al., 2006). As it can be seen, for the English-to-German task we did not use any of the previous enhancements.</Paragraph>
    <Paragraph position="2">  Important differences can be observed between the German-English and the rest of translation tasks. They result from the greater differences in word order present in this language pair (the German-English results are obtained under monotone decoding conditions). Also because the greater vocabulary of words of German, which increases sparseness in any task where German is envolved. As expected, differences in translation accuracy between Spanish-English and French-English are smaller.</Paragraph>
  </Section>
class="xml-element"></Paper>