A Path-based Transfer Model for Machine Translation

A Path-based Transfer Model for Machine Translation Dekang Lin Department of Computing Science University of Alberta Edmonton, Alberta, Canada lindek@cs.ualberta.ca Abstract We propose a path-based transfer model for machine translation. The model is trained with a word-aligned parallel corpus where the source language sentences are parsed. The training algorithm extracts a set of transfer rules and their probabilities from the training corpus. A rule translates a path in the source language dependency tree into a fragment in the target dependency tree. The problem of finding the most probable translation becomes a graph-theoretic problem of finding the minimum path covering of the source language dependency tree. 1 Introduction Given a source language sentence S, a statistical machine translation (SMT) model translates it by finding the target language sentence T such that the probability P(T|S) is maximized. In word-based models, such as IBM Model 1-5 (Brown et al 1993), the probability P(T|S) is decomposed into statistical parameters involving words. There have been many recent proposals to improve translation quality by decomposing P(T|S) into probabilities involving phrases. Phrase-based SMT approaches can be classified into two categories. One type of approach works with parse trees. In (Yamada&Knight 2001), for example, the translation model applies three operations (re-order, insert, and translate) to an English parse tree to produce its Chinese translation. A parallel corpus of English parse trees and Chinese sentences are used to obtain the probabilities of the operations. In the second type of phrase-based SMT models, phrases are defined as a block in a word aligned corpus such that words within the block are aligned with words inside the block (Och et al 1999, Marcu&Wong 2002). This definition will treat as phrases many word sequences that are not constituents in parse trees. This may look linguistically counter-intuitive. However, (Koehn et al 2003) found that it is actually harmful to restrict phrases to constituents in parse trees, because the restriction would cause the system to miss many reliable translations, such as the correspondence between “there is” in English and “es gibt” (“it gives”) in German. In this paper, we present a path-based transfer model for machine translation. The model is trained with a word-aligned parallel corpus where the source language side consists of dependency trees. The training algorithm extracts a set of paths from the dependency trees and determines the translations of the paths using the word alignments. The result of the training process is a set of rules for translating paths in the source language into tree fragments in the target language with certain probabilities. To translate a sentence, we first parse it and extract a set of paths from its dependency tree S. We then find a set of transfer rules that cover S and produce a set of tree fragments obtained to form a tree T* such that T*=argmaxT P(T|S). The output sentence can then simply be read off T*. In the remainder of the paper, we first define paths in dependency trees. We then describe an algorithm for learning transfer …

Content truncated to 500 words