Exploring the Use of Target-Language Information to Train the Part-of-Speech Tagger of Machine Translation Systems

Speeding Up Target-Language Driven Part-of-Speech Tagger Training for Machine Translation

“…As mentioned in section 3, the TL-driven training method needs a TL model to score the different translations τ (g i , s) of each SL text segment s. In this paper we have used a classical trigram language model like the one used in [3]. This language model was trained on a raw-text Catalan corpus with around 2 000 000 words.…”

“…The main disadvantage of the TL-driven method used to train HMM-based PoS taggers [3] is that the number of translations to perform for each SL text segment grows exponentially with the segment length. In this paper we have proposed and tested a new approach to speed up this training method by using a priori knowledge obtained in an unsupervised way from the SL.…”

“…However, when the resulting PoS tagger is to be embedded within a machine translation (MT) systems, the use of information not only from the source language (SL), but also from the target language (TL) has proven to give better results [3].…”

“…The TL-driven training algorithm [3] proceeds by translating every possible PoS tag sequence resulting from the disambiguation of the words in each SL text segment into the TL, and using a probabilistic TL model to estimate the likelihood of the translation corresponding to each possible disambiguation. The main disadvantage of this method is that the number of possible disambiguations to translate grows exponentially with the segment length.…”

“…It must be pointed out that the TL-driven training method described in [3], along with the pruning method proposed in this paper, have been implemented in the package name apertium-tagger-training-tools, and released under the GPL license. 1 The rest of the paper is organized as follows: Section 2 overviews the use of HMM for PoS tagging.…”