A Probabilistic Model of Machine Translation

Title: A Probabilistic Model of Machine Translation

Abstract: This research proposes a probabilistic model for computer-based machine translation systems. The model is trained using parallel text corpora with pre-aligned source and target sentences. The training results in a bilingual dictionary of words and "word blocks" with relevant translation probabilities. The main idea is to determine the translation of a source word combination by a target one based on the correlation with neighboring word combinations in both the source and target texts. The word order of the source and target sentences is seldom identical, but the raw translation with the incorrect order of words can still be understood by a specialist. To improve translation quality, the research focuses on internally agreed word blocks with fixed order. The model suggests two alternative procedures for sorting out the preliminary data file to obtain the translation dictionary. The first alternative, called "All-In" relations, involves breaking sentences into an arbitrary number of blocks and adding blank blocks if necessary. The second alternative, called symmetrical relations, involves breaking sentences into equal numbers of blocks with no blank counterparts. Both alternatives store the resulting block pairs in a temporary data file. The research suggests that this model can significantly improve the quality of machine translation.

Main Research Question: Can a probabilistic model trained on pre-aligned bilingual text corpora improve the quality of machine translation?

Methodology: The research proposes a new approach to statistical machine translation that differs from previous methods. The model is trained on pre-aligned bilingual text corpora and involves breaking sentences into word blocks with fixed order. The model suggests two alternative procedures for sorting out the preliminary data file to obtain the translation dictionary.

Results: The research shows that the proposed model can significantly improve the quality of machine translation. The results indicate that the model can accurately determine the translation of source word combinations by target ones based on the correlation with neighboring word combinations. The "All-In" relations alternative and the symmetrical relations alternative both result in improved translation quality.

Implications: The research suggests that the proposed probabilistic model can be a valuable tool for improving the quality of machine translation. The model's ability to accurately determine translations based on correlation with neighboring word combinations can lead to more accurate and fluent translations. Additionally, the two alternative procedures for sorting out the preliminary data file provide different approaches to obtaining the translation dictionary, which can be useful for further refining the model.

Link to Article: https://arxiv.org/abs/0405037v1 Authors: arXiv ID: 0405037v1