Abstract
A large body of translation texts has long been acknowledged to be a treasure of linguistic knowledge: it contains a wealth of lexical and syntactical information, among others, to be extracted. However, this information can only be utilized if we know which part of the text in one language corresponds to which part of the text in the translation. In other words, only if the translation texts have previously been aligned, preferably automatically. The alignment problem is mostly attacked from two directions: the one that based mainly on statistics (Brown et al., 1990; Brown et al., 1993; Church, 1993; Dagan, Church, and Gale, 1993; Chen, 1993), and the one that require some kind of lexical knowledge (Utsuro et al., 1994; Ker and Chang, 1997). But each has weakness of its own; dictionary based approaches usually fall short of the requirement of adaptability with regard to language evolution and domain change, while statistical approaches usually suffer from data sparseness, have simplistic repre-sentation and are hungry for computational resources. Particularly, the parameters estimation phase of a conventional statistical approaches usually involve fine-tuning of a huge and opaque mathematical model, which entails complicated computation, massive memory requirement and time consuming iterative processes. Owing to the reductionist mathematical modeling, it is also hard to incorporate new ideas and extend these systems: development become stale. In this paper, we will describe a variation of statistical approach to alignment. However, we view a sentence not as any random chain of morphemes, as previous statistical models do, but a meaningful, grammatical sequence which is susceptible to many constraints; zooming in on a finer resolution of structures within a pair of sentences, we saw certain distortion regularity between parallel texts emerges from these constraints. The regularity, which is overlooked by previous researches, was found to be highly useful in identifying connections for alignments. We cast the task of alignment into a pattern classification problem, and employed artificial neural networks and image processing techniques to take advantage of this regularity. The pattern classification framework with its generality had also enabled us to incorporate a language model and capture some local syntax of the sentences in the alignment program. The solution proposed here is linguistically neutral, i.e., it does not depend on the particular languages we choose. Other than a training data of modest size, it does not use any external resource such as bilingual lexicon, thesaurus, dictionary, etc. We focused on the discovering of overlooked information hidden in the data, and found that exploiting this new information is more cost effective than doing time consuming iteration to fine-tune the translation model Consequently, one pass through the training data had produced satisfactory result.