A Flexible Architecture for CAT Applications

A Flexible Architecture for CAT Applications Saˇsa Hasan, Shahram Khadivi, Richard Zens, and Hermann Ney Lehrstuhl f¨ur Informatik VI – Computer Science Department, RWTH Aachen University, D-52056 Aachen, Germany {hasan,khadivi,zens,ney}@cs.rwth-aachen.de Abstract We present an intuitive technical framework for making Computer Assisted Translation (CAT) adaptable and more suitable for rapid application devel- opment. The framework is a client-server-based architecture that uses an approach similar to “message passing”, a technique widely used in computer science. We define a “translation object”, a structure holding all necessary data, that is passed to server-like processes via sockets. This method can be easily enhanced in a modular manner where several recipients build a chain, one passing the processed object to the next one. We enhance a state-of-the-art phrase-based translation system with server and interactive generation capabilities and evaluate this prototype on different language pairs. 1 Introduction Computer Assisted Translation (CAT) aims at helping professional translators to faster translate texts from one language into an- other. The broad term covers many aspects, reaching from electronic dictionaries, termi- nology databases, automatic translation sys- tems and other modules, such as transla- tion memories. A crucial component is the machine translation system, as it imposes most of the computation and memory re- quirement constraints. Obviously, a sepa- ration of the translator’s environment and a dedicated translation server is intelligible (Och, Zens, & Ney, 2003). Generally, there might be additional com- ponents involved in the overall translation process, such as preprocessing, on-the-fly reranking and eventual postprocessing (e.g. truecasing). We present a straightforward framework that allows for several modules to be connected in series, employing a com- mon interface and defined data structures as input and output. Thus, the overall mainte- nance effort is facilitated. The idea is to use translation objects that hold all necessary information and pass them from one application to another. For flexibility reasons and ease of use, we choose TCP/IP sockets to accomplish this task. Socket modules are available in all major programming languages, such as C++ or Python. Each program therefore has to in- corporate only a small set of basic capa- bilities, i.e. receiving, parsing and sending the object, in order to be usable in the ap- plication chain. One major advantage is that many such modules can be provided by different research groups and easily set up for experimentation. By using TCP/IP, the servers even do not need to be in one in- tranet but can be located anywhere on the internet instead. Furthermore, to test our basic framework, we incorporated server-like capabilities and an interactive search mode in a state-of-the- art phrase-based machine translation (MT) system (Zens et al., 2005). The current per- formance is similar to an interactive machine translation (IMT) system based on align- ment templates. Related work in the CAT domain is re- ferred to in the next section. In Section 3, we review the theoretical framework for interac- tive machine translation being derived from a statistical MT viewpoint. In Section 4, we present a detailed overview on the techni- …

Content truncated to 500 words