Automatically trained dialogue systems

Abstract

A chatbot is a software system that is able to interact with a human via natural language conversation (Bayan Abu Shawar, 2007). The Loebner Price has provided an environment in which the chatbots ability to fool a judge is put to test. One common drawback in most chatbot systems is the manual creation of the corpus from which it can draw answers to a user's utterance. In this project we investigated two techniques that try to automatically tram these systems by using movie scripts. The first approach taken tries to convert the movie scripts into the AIML structure (used quite often for the development of chatbots) by making use of text categorisation techniques to create a series of networks for fast access. The second approach use information retrieval techniques to automatically train the dialogue system. A system that makes use of this corpus generated from the movie scripts is constructed. In addition an online Question Answering system and leading chatbot were added to the system. These sources are governed by a dialogue manager to produce the best output. The system is put to test by comparing it against Alice bot (one of the most successful chatbots ). Our conclusion was that the use of movie scripts to automatically train a chatbot/dialogue system is a successful alternative to the manual creation of the corpus.