Arabic goal-oriented conversational agents using semantic similarity techniques

Conversational agents (CAs) are computer programs used to interact with humans in conversation. Goal-Oriented Conversational agents (GO-CAs) are programs that interact with humans to serve a specific domain of interest; its’ importance has increased recently and covered fields of technology, sciences and marketing. There are several types of CAs used in the industry, some of them are simple with limited usage, others are sophisticated. Generally, most CAs were to serve the English language speakers, a few were built for the Arabic language, this is due to the complexity of the Arabic language, lack of researchers in both linguistic and computing. This thesis covered two types of GO-CAs. The first is the traditional pattern matching goal oriented CA (PMGO-CA), and the other is the semantic goal oriented CA (SGO-CA). Pattern matching conversational agents (PMGO-CA) techniques are widely used in industry due to their flexibility and high performance. However, they are labour intensive, difficult to maintain or update, and need continuous housekeeping to manage users’ utterances (especially when instructions or knowledge changes). In addition to that they lack for any machine intelligence. Semantic conversational agents (SGO-CA) techniques utilises humanly constructed knowledge bases such as WordNet to measure word and sentence similarity. Such measurement witnessed many researches for the English language, and very little for the Arabic language. In this thesis, the researcher developed a novelty of a new methodology for the Arabic conversational agents (using both Pattern Matching and Semantic CAs), starting from scripting, knowledge engineering, architecture, implementation and evaluation. New tools to measure the word and sentence similarity were also constructed. To test performance of those CAs, a domain representing the Iraqi passport services was built. Both CAs were evaluated and tested by domain experts using special evaluation metrics. The evaluation showed very promising results, and the viability of the system for real life.