Hello! I have seen many examples of creating a chat bot based on classification. There is an array of intents, where inside there is a tag, a question and an answer. And the neural network tries to predict which tag is the most suitable, and then the script takes a random answer from the array.
This is interesting, but as a result, the neural network responds with templates and it is very difficult to replenish the array of intents by predicting what the person will ask the chat bot.
Is it possible in TensorFlow JS or Lite to train a model on an array of answers and questions that will not be stored next to the model in the form of JSON or something like that, and the neural network will learn and remember all questions and answers, words, sentences. And by loading this model, the neural network will be able to formulate its proposals based on the trained data, answering the question that the person asks. For example, there is an array of conversations between a client and a manager, and having received these dialogues, the neural network will try to answer the client, as if it were a manager who knows his field.
It turns out that the neural network does not use templates, but generates its answers based on the knowledge obtained from the question-answer during training.
Are there such possibilities? If so, what are the lessons or maybe a video on how to implement this?