Systems and methods for training a machine learning based natural language processor to process specialized language terms. A training corpus of labeled specialized language terms associated with a specialized subject area is assembled. A generated database query corresponding to a question comprising at least one specialized language term within the set of labeled specialized language terms is created with a machine learning based natural language processing process. A cost function is determined based on differences between the generated database query and contents in a knowledge graph mapping terms describing characteristics of entities associated with the specialized subject area. The natural language processing process is refined based on feeding back the cost function to create a refined natural language process. Received queries are processed with the refined natural language process to create database queries directed to answering questions associated with the received query, and the answer is presented.