Improving Service Chatbot Using Semantic Based Short-Term Memory

Chatbots have become integral across various industries in Thailand, spanning online product ordering, call centers, customer feedback collection, and service robots. However, effective dialog design is essential for their optimal performance. Currently, chatbots heavily rely on pattern matching techniques [1] for constructing conversational sentences, which can lead to errors when encountering unfamiliar patterns or semantically similar words. In those cases, the chatbot may terminate the conversation or necessitate human intervention to proceed. Additionally, lack of prior conversation's information presents a hurdle, especially for service robots or virtual assistants. For instance, when a customer seeks to cancel an order, the chatbot requires access to precise order information. To overcome this problem, Long-term memory in Chatbot has been introduced in [2] and [3]. However, Long-term memory in Chatbot is not effectively used in certain scenarios. For example, service robot serving in the restaurant, the information such as number of seats to be served, list of ordering food, additional requested utensils are only used while the customers are in the restaurant and can be terminated afterward. In this case the long-term memory usage may not be indispensable. We proposed SEBASS or a Semantic-Based Short-Term Memory service chatbot to cope with this problem. We performed experiments to compare the effectiveness between the traditional chatbot (BASELINE) which has no memory and the proposed semantic-based short-term memory chatbot (SEBASS). Restaurant service scenarios in the restaurant such as food ordering and food cancelation are used as test scenarios. The results showed that SEBASS outperforms BASELINE in terms of interaction success rate and incorrect response rate. SEBASS has an interaction success rate equal to 100% compared to 96.71% of BASELINE and SEBASS has an incorrect response rate equal to 0% compared to 0.76% of BASELINE.