Computer-generated Dialogues through Sequential Pattern Mining
MetadataShow full item record
In this thesis I attempt to answer the following question: How does generating artificial, goal-oriented dialogue using automated data mining weigh up against a manual approach to this problem? I wrote the Pet Shop Game, a web application that anonymously couples players and allows them to engage in a dialogue game set in a pet shop scenario. Players can express themselves in a controlled natural language and perform relevant physical actions. The dialogue data is then mined for sequential patterns using the GSP algorithm, which also provides a taxonomy parameter that is used in this case to not only find patterns at the utterance level but also on the level of speech acts. Sequential patterns can be translated into rules suitable for use by an AI program to replace a human player in the Pet Shop Game. These data mined rules, though arguably of the same structure as rules used by hand-written chatbots, are weaker and find their strength in numbers. That is, many data mined rules combine to come to the right conclusion, whereas hand-written rules are typically much stronger. I suggest a comparative study of serious implementations of both approaches should be made in order to arrive at a more definitive judgement.