A semantic-based technique for question lassification in question answering systems — A hybrid approach

Question classification plays a substantial role in Question Answering systems. To obtain a semantically rich and syntactically sound question classifier, we present in this paper a semantic-based hybrid approach that explores the meaning underlying the question and seeks to understand the linguistic denotation of the question by constructing a lexically minimalistic syntactic labelled graph from it Questions are classified into one of the six major and fifty finer classes. The system proposes a training method that learns a set of graph traversal rules to detect a question's focus. A standard set containing diverse classes of benchmark questions is used for the training purpose. The system proposes using a semantic memory to store a set of commonly used and unambiguous words that often occur with specific graph traversal rules. A modified technique of the Word Sense Disambiguation using WordNet detects the contextual meaning of a question's focus and maps it to a finer class. A substantial improvement achieved using the approach over other similar systems using similar benchmark data shows that the current approach can be used as a syntactically and semantically rich model in the area of question answering systems.