A Neuro-Proximal Hybrid Framework Optimizing Multi-Intent Conversational AI via MiniLM Embeddings

Abstract

The global conversational Artificial Intelligence (AI) market is on a trajectory to reach USD 32 billion by 2030, with chatbots expected to facilitate more than 80% of customer engagements. However, the scalability of these services is currently throttled by manual intent annotation and query classification, which are often inconsistent and labour-intensive. To resolve these bottlenecks, this research introduces the Neuro-Proximal Hybrid (NPH) framework. This system leverages a multi-intent and categoryannotated Customer Support Bitext dataset. The methodology begins with Natural Language Processing (NLP) preprocessing and Exploratory Data Analysis (EDA) to normalize, tokenize, and visualize data distributions. A Miniature Language Model (MiniLM) is then implemented to provide lightweight yet contextually rich feature extraction. To mitigate dataset imbalances, the Synthetic Minority Oversampling Technique (SMOTE) is utilized to generate artificial examples for minority classes. Moving beyond traditional models like the Decision Tree Classifier (DTC), KNN model, or Naïve Bayes Classifier (NBC), the NPH pipeline integrates Deep Neural Network (DNN) feature selection with KNN to optimize classification. The system targets two bivariate outputs such as Intent and Category to deepen the contextual interpretation of user queries. By integrating this model into a live chatbot interface, the framework enables automated, real-time response generation. The NPH approach significantly improves accuracy, reduces human error in annotation, and elevates customer satisfaction through superior multi-intent comprehension.