| Title |
Recursive Dialog Summarization for Long-Term Memory Transformers |
| DOI |
https://doi.org/10.5573/ieie.2024.61.9.92 |
| Keywords |
Long-term memory; Natural language processing; Text summarization |
| Abstract |
Recently, AI chatbots using Transformer[2]-based language models such as ChatGPT[1] have been attracting attention. A significant challenge faced by these models is their inability to maintain coherent context throughout a conversation, often losing initial dialogue information. To address this problem, we introduce a novel algorithm called Recursive Dialog Summarization (RDS). RDS dynamically condenses the length of the conversation history, replacing it with a summarized version for the language model. Experiments demonstrate that RDS not only enhances the response accuracy of the language models but also effectively reduces the number of input tokens, leading to faster latency. Through a real-world chatbot, we show that dialog summarization can maintain users' personal information in long conversations. |