Unsupervised Weighting of Transfer Rules in Rule-Based Machine Translation using Maximum-Entropy Approach

In this paper we present an unsupervised method for learning a model to distinguish between ambiguous selection of structural transfer rules in a rule-based machine translation (MT) system. In rule-based MT systems, transfer rules are the component responsible for converting source language morphological and syntactic structures to target language structures. These transfer rules function by matching a source language pattern of lexical items and applying a sequence of actions. There can, however, be more than one potential sequence of actions for each source language pattern. Our model consists of a set of maximum entropy (or logistic regression) classifiers, one trained for each source language pattern, which select the highest probability sequence of rules for a given sequence of patterns. We perform experiments on the Kazakh - Turkish language pair - a low-resource pair of morphologically-rich languages - and compare our model to two reference MT systems, a rule-based system where transfer rules are applied in a left-to-right longest match manner and to a state-of-the-art system based on the neural encoder-decoder architecture. Our system outforms both of these reference systems in three widely used metrics for machine translation evaluation.