Personalized Chat and Dialog
- EBDM Framework
- Personalized Knowledge Base and Relevance Score for Chatting
- Contact
- Jeesoo Bang
Recently natural language processing technology makes its way into our daily life. Many people has great interest on the human-friendly interface, and they are amazed at how can a machine have such an ‘ability’. Among those, chatting system is one of the hottest technology receiving great attention. Most users expect that their machine can react and respond to their speech. This ability makes the machine looks smart and more attractive one. It would be even more attractive if it can talk about wide range of subjects and can adapt to the user. Users can feel as if it has better performance, then the user satisfaction could be increased.
Responding to various kinds of user utterances is one of the important ability to the chatting system. Moreover, the ability should be extendible without difficulty. EBDM Framework is suitable for this kind of system. It is not that difficult to constructing (user utterance, system response) pairs or mining them from large data sources. Only limited amount of human annotations are required. The system will search the most relevant example among them and then give back corresponding system response. Relevance score need to be defined to find the appropriate example. PoS tags and dependencies information is used to measure the semantic similarity of two sentences, as the classical IR problems did.
Personalization of a chatting system can be achieved by remembering the user interests and applying it to the system response. Those kinds of information should be collected during the natural conversation between the user and the system. Knowledge triples are good container for those information and they are extracted from every user utterances. The information received from the user may need to remains for a while, and some should be remembered for long time, but not all of them. To distinguish the long-term memory and short-term memory, I will introduce memory model which includes interest propagation, forgetting, refreshing. By adopting memory model, we can distinguish the important information for the user from big pile of useless information. After we construct user memory, we can reflect it to the relevance score in EBDM chatting system. As a result, the system can refer the user’s interest more frequently, or can talk in user’s style. The ultimate goal is to make the user feel subtle changes as he or she converses with the system.