Learning of Translation by the Inductive Learning System LS/1

Abstract

LS/1 is a domain independent inductive learning system, which repeats Question-Response-Answer (Q - R - A) interactions with a teacher, and learns the structure of the relation between questions and answers. Each message (Q, R or A) is a sequence of words. LS/1 can be applied to the task of learning English-Japanese translation without any changes in the algorithm of the system by using a training sequence which consists of English sentences as questions and their Japanese translations as answers. In a learning experiment on an easy English-Japanese translation sequence, LS/1 effectively acquires knowledge for translation, which includes word-to-word correspondence and grammatical knowledge. The important point we stress here is that the successful result in learning of translation has been attained by an inductive learning system whose learning algorithm and knowledge representation system are constructed independently of concepts and terms related to translation, and whose initial knowledge involves no domain specific, syntactic or semantic knowledge. The knowledge representation system of LS/1 is called a label net, which has an expressive power to represent relational data bases, concept hierarchy and certainty factors of rules. We define basic concepts of label nets, such as deductivity, explainability and applicability. The response generation procedure of LS/1 utilizes a best first search algorithm to find the most plausible solution sentence. Label nets are closely related to the axiom sets of Horn clauses. Concepts and algorithms similar to the ones mentioned above can also be discussed in the Horn clause framework. The label net of LS/1 is more appropriate than Horn clauses when we construct learning systems which can discover unknown structures in sentence data. The learning algorithm of LS/1 includes generalization and merging algorithms of rules, generalization by identification of two predicates, and changing certainty factors in rules. Such learning algorithms are controlled by some conservative heuristics in order to avoid combinatorial explosions. This paper suggests that LS/1 is an effective research framework for the learning of translation as well as for the learning in other question-answering processes, and that research of translation learning in our framework is very useful in extending the domain independent theory of inductive learning.