Scalable Multi-Intent and Category Detection in Customer Support Using Lightweight Embeddings and Hybrid Feature Selection

Abstract

The global conversational Artificial Intelligence (AI) market is expected to reach USD 32 billion by 2030, with more than 80% of customer interactions managed by chatbots. Despite this growth, manual intent annotation and query classification remain time-intensive and inconsistent, limiting scalability in customer support systems. To address these challenges, this study proposes an advanced Natural Language Processing (NLP) framework built on a Customer Support Bitext dataset annotated with multiple intents and categories. The approach begins with NLP preprocessing and Exploratory Data Analysis (EDA) to clean, normalize, tokenize, and examine data patterns. For feature extraction, the Miniature Language Model (MiniLM) is employed to generate lightweight yet contextually rich embeddings. To handle class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is applied, creating synthetic samples for underrepresented classes and improving model fairness. Unlike traditional classifiers such as Decision Trees, K-Nearest Neighbors (KNN), and Naïve Bayes, the proposed framework integrates Deep Neural Network (DNN)-based feature selection with KNN for enhanced classification performance. The system is designed to predict two correlated outputs—Intent and Category—enabling deeper contextual understanding of customer queries. Finally, the model is deployed within a chatbot interface to support real-time intent detection and automated responses. This framework improves classification accuracy, reduces annotation inconsistencies, and enhances customer satisfaction through effective multi-intent recognition. Overall, the study demonstrates a scalable and efficient solution for advancing conversational AI in customer support applications.