Data management in training AI chatbot with personality based on granular computing
Piotr Podolski, Tomasz Ludziejewski, Hung Son Nguyen · 2022 IEEE International Conference on Big Data (Big Data) · 2022
In this article we present a set of data handling techniques and improvements in training a neural conversational model on big data. We approach this problem as a granular computing problem in which we focus on mini-batches creation as a problem of granules creation and relations between them to optimize the performance of the model. We also provide an empirical summary of influence on training time and model performance of different mini-batch creation strategies for training conversational neural model with personality. Additionally we present our own uniform length batching approach to speed up the training. We claim that with a mindful data preparation and use of data during training we are able to reduce compute time and infrastructure costs, thus allowing to train models on bigger data at reasonable time. The presented model is a chat-bot with topic awareness for a mobile service provider, which was trained on a large dataset of historical utterance data, that contained diverse subtopics and products. By applying data optimization techniques for handling and preparing batches of data, we can reduce the time of training and save a lot of time and compute costs without loosing performance of the model.