Architectures of “toy” systems for teaching machine translation

Architectures of "toy" systems for teaching Machine Translation Walther v. Hahn, Cristina Vertan University of Hamburg, Faculty of CS, Natural Language Systems Dept. vhahn, cri@nats.informatik.uni-hamburg.de Abstract This paper addresses the advantages of practical academic teaching of machine translation by implementations of „toy" systems. This is the result of experience from several semesters with different types of courses and different categories of students. In addition to describing two possible architectures for such educational toy systems, we will also discuss how to overcome misconceptions about MT and the evaluation both of the achieved systems and the learning success. 1. Introduction Machine Translation (MT) is an important sub-field both of Computational Linguistics and Natural Language Processing. Therefore academic education in MT addresses students in linguistics and computer science. Usually, according to the background of the students, courses are given separately to these groups and with different methods: theoretical aspects and demonstration of tools for the linguists (Somers, 2000) on one hand, implementation of clear defined algorithms for the computer scientists on the other hand. Both ways are successful when we aim at illustrating specific methods only. The disadvantage of this limited aim is that students will not be able to evaluate, configure and design MT system. Accordingly, they cannot answer questions like "How will the system react with huge amounts of data or maximal throughput?", "How much are supported changes of the domain or changes of languages?", "What about maintenance by users?" Or, more complicated: "What type of architecture is optimal for my required functionality?", "What kind of grammar is more suitable?" The alternative, however, to implement a real MT-system in a course is not feasible, due to lack of time and missing background knowledge of the students. Very often they are facing the field for the first time. A solution in between may be the implementation of "toy systems", with limited language resources and limited functionality. In this paper we will two discuss examples of such toy systems, implemented by our students in two different types of courses. In section 2 we will describe the context of the courses: The course structures at Hamburg University, the language background of the students and the preparatory material for students. Section 3 presents the architecture and functionality of the two systems. In section 4 we will give details of the students' evaluation of their systems and learning success, followed by a discussion of the students' misconceptions about MT at the beginning of the course. 2. Organisational aspects 2.1. Type of courses and participants To understand the following paragraphs, it is necessary to give some details about the structure of courses at Hamburg University (HHU). We offer three types of courses: lectures (accompanied by weekly exercises) seminars and projects. Lectures normally address central issues of 69 Computer Science, like Object Oriented Programming, Computer Architectures, Formal Languages etc. Specific fields are often taught in the frame of "seminars" and projects. Usually, seminars, after an introduction by the professor(s), consist of student presentations of a particular topic, according …

Content truncated to 500 words