Chatbot Model Development and Training: Bidirectional LSTMs, Word2Vec Embeddings, and Context-Aware Response Design

Introduction

This paper walks through the construction of a machine learning chatbot for clinical data insights. It covers the training dataset (370+ labeled examples across 20 intent categories), a 9-step preprocessing pipeline with lemmatization and TF-IDF analysis, Word2Vec embeddings, and a Bidirectional LSTM architecture chosen for its ability to capture contextual dependencies in both directions. The model achieved 97.3% validation accuracy with a context-aware response system integrating VADER sentiment analysis, spaCy NER, and a 3-turn dialogue state tracker over real patient demographics, hospital admissions, ICD-10 diagnoses, and lab results.