MaSMT: A Multi-agent System Development Framework for English-Sinhala Machine Translation B. Hettige #1 , A. S. Karunananda *2 , G. Rzevski *3 # Department of Statistics and Computer Science, University of Sri Jayewardenepura, Nugegoda, Sri Lanka 1 budditha@dscs.sjp.ac.lk * Faculty of Information Technology, University of Moratuwa, Sri Lanka, 2 asoka@itfac.mrt.ac.lk 3 rzevski@gmail.com Abstract—Multi-agent System gives high quality solutions through communication, negotiation and coordination among agents. Agents are small self-contained computational objects capable of exchanging messages among themselves. Number of general purpose toolkits and frameworks are available to develop Multi-agent systems for modelling complex real world problems. However, none of them has been specialized for the area of natural language processing specially machine translation. MaSMT is a java based multi-agent system development framework, especially designed for development of English to Sinhala machine translation system. MaSMT provides two types of agents, namely ordinary agents and manager agents. A manager agent consists of number of ordinary agents within its control. Further, manager agents can directly communicate with other manager agents and each and every ordinary agent in the swarm is assigned to a particular manager agent. An ordinary agent in a swarm can directly communicate only with the agents in its own swarm and its manager agent. The framework primarily implements object-object communication, XML-based data passing and MySQL database connectivity to use agents’ ontology and message passing. Agent communication in the framework has been implemented to comply with FIFA-ACL specification. MaSMT framework is used to develop English to Sinhala machine translation system and Word Reader which is capable of analysing a given English word. Experimental result shows that, MaSMT framework can be used to develop Natural Language Processing applications successfully. Keywords— Machine Translation, Multi-agent Systems. I. INTRODUCTION Multi-agent System (MAS) technology has emerged as a new software paradigm, which exploits the power of message passing as the key strategy for problem solving. Communication, negotiation and coordination among agents produce high quality solutions that cannot be generated by a single agent in its individual capacity. Nowadays hundreds of well-established general purpose toolkits and frameworks are available for the development of Multi-agent Systems. Among others, JADE [1], Jason [2], AgentBuilder [3] and SeSAm [4] are the standard Multi-agent System development frameworks. JADE (Java Agent DEvelopment Framework) is a software framework fully implemented in Java language. JADE framework provides supporting GUI tools for debugging and deployment phases in multi agent developments. Jason is an interpreter for an extended version of AgentSpeak [5]. AgentBuilder is an integrated software development tool that allows software developers to build agents quickly and easily without sound knowledge of Multi-agent technology. SeSAm (Shell for Simulated Agent Systems) is another framework that provides a generic environment for modelling and experimenting with agent-based simulation. Further, agent development framework saves developers‟ time and provides standardization of the multi-agent system development [6]. Many of these frameworks are especially designed to develop general purpose applications, machine learning and simulations of complex systems. However, these existing frameworks do not support the distinct requirements of Machine Translations, coming under area of Natural Language Processing. Machine Translation System is computer software that translates text or voice from one natural language into another with or without human assistance [6]. Machine Translation system produces translation through three major steps including analysis of source language text, translation and generation of the texts in the target language [7]. These sub systems are required to handle morphology, syntax and semantic aspects of two languages. Words in a language are employed as the building block of natural language understanding. This is valid for people who read a sentence word by word or otherwise by locating selected words such as nouns and verbs. Consequently, meaning of a sentence is determined by the interaction among words, which draw from all aspects of morphology, syntax and semantic, as appropriate. “Words in a sentence as agents”, [8] is the philosophy behind to design English to Sinhala machine translation system. These word agents pass messages among them within and across different level of analysis. As such, this machine translation approach is different from existing ones that sequentially define linguistic aspects such as morphological, syntax and semantics analysis. Further, English to Sinhala Machine Translation system is capable of processing English Morphological analysis, English Syntax analysis, English to Sinhala Semantic level translation, Sinhala Morphological generation and Sinhala syntax generation through the Multi-agent approach [9].