This paper presents a machine learning approach to question classification, aiming to categorize questions into fine-grained semantic classes that provide constraints on potential answers. The authors develop a hierarchical classifier guided by a layered semantic hierarchy of answer types, which is evaluated on a large collection of free-form questions from the TREC 10 competition. The classifier is trained using the SNoW learning architecture and is evaluated on both coarse and fine classes. The results show that the hierarchical classifier achieves high accuracy (over 90%) despite the large number of labels (50). The study also highlights the importance of semantic features in achieving good performance, as local features alone are insufficient. The experimental results demonstrate the effectiveness of the learned classifier in question classification, with a precision of 98.80% for coarse classes and 95% for fine classes. The paper concludes by discussing future directions, including deeper semantic analysis and automated feature generation.This paper presents a machine learning approach to question classification, aiming to categorize questions into fine-grained semantic classes that provide constraints on potential answers. The authors develop a hierarchical classifier guided by a layered semantic hierarchy of answer types, which is evaluated on a large collection of free-form questions from the TREC 10 competition. The classifier is trained using the SNoW learning architecture and is evaluated on both coarse and fine classes. The results show that the hierarchical classifier achieves high accuracy (over 90%) despite the large number of labels (50). The study also highlights the importance of semantic features in achieving good performance, as local features alone are insufficient. The experimental results demonstrate the effectiveness of the learned classifier in question classification, with a precision of 98.80% for coarse classes and 95% for fine classes. The paper concludes by discussing future directions, including deeper semantic analysis and automated feature generation.