|
User Experience of Natural Language Interaction with a Generative Artificial Intelligence System Ştefan Trăuşan-Matu |
67-84 |
|
Enhanced Human-Machine Conversations by Long-Term Memory and LLMs Dan-Constantin Dumitriu, Dragoş-Florin Sburlan |
85-102 |
1 University Politehnica of Bucharest
313 Splaiul Independentei, Bucharest, Romania
2 Institutul de Cercetări în Inteligenţa Artificială
Calea 13 Septembrie nr. 13, Bucureşti
3 Academy of Romanian Scientists
Splaiul Indpendentei 54, Bucharest, Romania
Abstract. Due to the wide use of chatbots nowadays, the investigation of user experience (UX) when interacting with these artificial intelligence (AI) artifacts is essential. Even if there already are some works with similar aims, they do not study UX in the perspective of some specific human features, such as dialogism, pragmatics, and style, which should be considered in the subdomain of AI-oriented to the human factor. The paper introduces the existent UX laws, which are rather directed towards visual, direct manipulation interfaces, with the aim of emphasizing the differences introduced to the UX by the human-chatbot interaction, using natural language. The originality of this paper is the consideration of style and the polyphonic model in analyzing human-chatbot interaction.
Keywords: user experience; generative artificial intelligence; ChatGPT; natural language processing; human-chatbot interaction; human factors, style, polyphonic model
Cite this paper as:
Trăuşan-Matu, S. User Experience of Natural Language Interaction with a Generative Artificial Intelligence System.
International Journal of User-System Interaction 16(3),
67-84, 2023.
Ovidius University of Constanţa, Faculty of Mathematics and Computer Science
124 Mamaia Blvd., 900527, Constanta, Romania
Abstract. In this article, we explore the quality of the conversation between a human and a machine, the depth, relevance, and the “personalization” as opposed to an answer extracted from a pre-trained body of knowledge. To this end, we focus on the importance of memory augmenting a LLM, by designing a system whose answers in conversation retrieve information from learned details. LLM systems have indeed the concept of context that achieves this goal, however, the LLM concept of context is a component that is ephemeral and limited in size. In this article, we explore a Long-Term Memory (LTM) solution and a way to extract relevant details precisely. The concepts and information stored in this memory are built upon the prior conversations between the human and the system, illustrating the parent-child education paradigm. Although there is no unanimity regarding a standardized test that measures or quantifies how personal an artificial system is, our results empirically prove the method's feasibility (even if the proposed method has its limitations, the richness of possible extensions is worth studying).
Keywords: Personalized Human-Machine Conversation, Personalized AI, Personalized LLM, Long-Term Memory LLM, LLM Memory, LLM Education
Cite this paper as:
Dumitriu, D.-C., Sburlan, D.-F. Enhanced Human-Machine Conversations by Long-Term Memory and LLMs.
International Journal of User-System Interaction 16(3),
85-102, 2023.