Building And Improving A Seq2seq-Based Chatbot Model

Abstract

Chatbots, or conversation agents, are software systems designed to generate responses based on given inputs, simulating conversations in text or voice. These systems are designed to mimic human-like interactions, providing users with a friendly and engaging experience. They have found being used in customer service, product recommendation systems in online stores, and personal assistance. Most of these chatbots are rule-based, providing responses based on predefined templates or rules that are simple to build and can handle most situations before a human assistant takes over. However, with the advent of deep learning, there have been numerous attempts to create more advanced conversation agents. These agents are based on generative models, which allow them to generate new responses based on existing human conversations. This means that there is little to no need for humans to create new rules for the chatbot to answer new questions. A promising approach in this domain is the use of Sequence-to-Sequence models, which have proven to provide a high level of coherence and relevance in responses. These models are based on Recurrent Neural Networks (RNNs), which have proven effective in natural language processing tasks. There are several types of RNNs, such as Long Short-Term Memory Networks (LSTMs) and Gated Recurrent Units (GRUs). Both approaches have been implemented in various ways, demonstrating how they can be run and managed to provide responses and engage in short test conversations. However, it is not clear whether these approaches are effective, as no tests have been conducted to collect results and evaluate the approaches, despite the presence of instructions on how to collect result data. Therefore, this work will attempt to construct chatbots based on LSTMs and GRUs, which are part of the Seq2Seq model family, to test and evaluate these existing approaches. These approaches are based on well-established principles and theories, and the results could be either good or bad. Comparisons have also been made with other research papers to get an idea of their effectiveness. However, direct comparisons are impossible as only results are shown, with no mention of implementation methods, frameworks, or how the results were obtained and visualized. In the best case, these papers only provide the formulas from well-established theories and principles that everyone knows, but no details on the implementation method are provided. Some may use different implementations and datasets.