<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2061">
  <Title>Integration of Speech to Computer-Assisted Translation Using Finite-State Automata</Title>
  <Section position="7" start_page="472" end_page="472" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have studied different approaches to integrate MT with ASR models, mainly using finite-state automata. We have proposed three types of transducers to rescore the ASR word graphs: lexiconbased, phrase-based and fertility-based transducers. All improvements of the combined models are statistically significant at the 99% level with respect to the baseline system, i.e. ASR only.</Paragraph>
    <Paragraph position="1"> In general, N-best rescoring is a simplification of word graph rescoring. As the size of N-best list is increased, the results obtained by N-best list rescoring approach the results of the word graph rescoring. But we should consider that the statement is correct when we use exactly the same model and the same implementation to rescore the N-best list and word graph. Figure 1 shows the effect of the N-best list size on the recognition WER of the evaluation set. As we expected, the recognition results of N-best rescoring improve as N becomes larger, until the point that the recognition result converges to its optimum value.</Paragraph>
    <Paragraph position="2"> As shown in Figure 1, we should not expect that word graph rescoring methods outperform the N-best rescoring method, when the size of N-best lists are large enough. In Table 2, the recognition results are calculated using a large enough size for N-best lists, a maximum of 5,000 per sentence, which results in the average of 1738 hypotheses  ent N-best sizes on the evaluation set.</Paragraph>
    <Paragraph position="3"> per sentence. An advantage of the word graph rescoring is the confidence of achieving the best possible results based on a given rescoring model.</Paragraph>
    <Paragraph position="4"> The word graph rescoring methods presented in this paper improve the baseline ASR system with statistical significance. The results are competitive with the best results of N-best rescoring. For the simple models like IBM-1, the transducer-based integration generates similar or better results than N-best rescoring approach. For the more complex translation models, IBM-3 to IBM-5, the N-best rescoring produces better results than the transducer-based approach, especially for IBM5. The main reason is due to exact estimation of IBM-5 model scores on the N-best list, while the transducer-based implementation of IBM-3 to IBM-5 is not exact and simplified. However, we observe that the fertility-based transducer which can be considered as a simplified version of IBM-3 to IBM-5 models can still obtain good results, especially if we compare the results on the evaluation set.</Paragraph>
  </Section>
class="xml-element"></Paper>