# Profitable Trade-Off Between Memory and Performance In Multi-Domain Chatbot Architectures

**Authors:** D Emre Tasar, Sukru Ozan, M Fatih Akca, Oguzhan Olmez, Semih Gulum, Secilay Kutay, Ceren Belhan  
**arXiv:** [2111.03963](https://arxiv.org/abs/2111.03963)  
**Date:** 2021-11-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.03963v1/

## Key points

- Study focuses on text classification problem in NLP within chatbot architectures
- Goal is to determine the class of a given text
- BERT, a commonly used method in NLP, is employed
- Single model used for multiple classification problems to reduce server load
- Masking method applied during estimation of single BERT model for classification across different subjects
- Three separate datasets covering different fields and closely related classification problems included in study
- Dataset consists of five classification problems with 154 classes
- Performance and server space compared between BERT model containing all classification problems and other BERT models trained specifically for each problem
- Research explores effective solutions for text classification within chatbot architectures using BERT models while reducing server load and achieving accurate classifications
- Insights provided on optimizing memory usage and performance trade-offs in multi-domain chatbot architectures

## Summary

This study focuses on the text classification problem in natural language processing (NLP) within chatbot architectures. The goal is to determine the class to which a given text belongs. To achieve this, Bidirectional Encoder Representations for Transformers (BERT), a commonly used method in NLP, is employed. A single model for multiple classification problems is used to reduce server load caused by using multiple models. This is done by applying a masking method during the estimation of a single BERT model created for classification across different subjects. Three separate datasets covering different fields and including closely related classification problems are also included in the study. The dataset consists of five classification problems with 154 classes. The performance and server space occupied by a BERT model containing all classification problems are compared with those of other BERT models trained specifically for each problem. This research explores how to effectively solve text classification problems within chatbot architectures using BERT models while alleviating server load and achieving accurate classifications. Insights into optimizing memory usage and performance trade-offs in multi-domain chatbot architectures are provided as well.