In this work, an idea for an 'Emotional-Chat' project based on the field of NLP is included.
This repository contains the code for "Emotional-Chat" introduced in this repo.
The code is built on Pytorch.
After COVID-19, modern people's psychological difficulties such as depression, anxiety, and loneliness are emerging as important social problems. This 'Emotional Chat' is designed to reduce the cost and time burden of users experiencing psychological difficulties through a counseling chat bot so that they can more easily access psychological counseling. 'Emotional Chat' is a conversation with users and Q&A. Through this, it extracts and analyzes texts that contain emotions and psychology, and derives users' psychological problems and emotions. Through this process, ‘Emotional Chat’ communicates with users, provides psychological comfort, and further suggests simple psychological prescriptions. ‘Emotional Chat’ is a natural language processing-based learning model and learns and processes through scripts for emotional AI corpus data.
In modern society, the ability of machines to understand and respond to human emotions is becoming increasingly important. This is particularly evident in interactive Artificial Intelligence (AI) chatbots, sentiment analysis, and personalized recommendation systems. With the evolution of the era, research is actively being conducted to make machines converse naturally like humans, comprehend people's emotional states, and exhibit appropriate emotional responses.
In this context, our research focuses on three primary Korean language processing AI models: keT5, koBart, and koBert. Each model has shown remarkable results in the Natural Language Processing (NLP) field recently, and the goal of this research is to compare and analyze the emotional conversation generation capabilities of these models.
Through this, it will be helpful in understanding how each model comprehends text and how they generate emotional responses.
The experimental data for this study consists of 270,000 sentences of conversational sentiment data, utilizing files containing 50,000 json-formatted sentimental conversations. Each model was trained on the same dataset, and their performances were compared and analyzed based on the same evaluation criteria. The findings of this research provide significant insights for the development of sentiment chatbots and automatic sentiment analysis systems and a deeper understanding of the AI models' capabilities in emotional comprehension and response generation.
The focus is not just on how well each model performs utterance, but on which among keT5, koBart, and koBert performs better with the same preprocessing.