As opposed to factoid questions, questions posed in a closed domain are typically more open-ended. People can ask for specific properties, procedures or conditions and require longer and more complex answers. As a result, detailed understanding of the question and the corpus texts is required for answering such questions. In this paper, we present a unification-based algorithm for measuring syntactic and semantic similarity of a question to candidate sentences extracted by information retrieval. The algorithm first applies strict linguistic constraints in order to identify potentially similar sentences to the question, then uses a statistical method to measure the similarity of the question’s subject and object to text chunks in each of these sentences. The algorithm has been evaluated on a closed domain in telecommunications and on the TREC 2003, 2004 and 2005 questions about the AQUAINT corpus for comparison. The evaluation shows a precision of ¢¤£¦ ¥ §¤ ¨ in our telecommunications domain, and ©¤§¤ ¨ on the TREC noncopulative questions. This confirms our hypothesis of the applicability of deep syntactic analysis for closed domain QA. 1

Abstract. Many exciting things happened since I left the ExtrAns project at the University of Zurich. In this paper I present a very brief journal of the research that derived from my work at Zurich. My work combined a study of several representations of questions and answer sentences with the development of procedures to find the answer. The culmination of this work was the definition of the Logical Graphs (LGs), which are graph representations derived from ExtrAns ’ original Minimal Logical Forms (MLFs), and the development of a method for the automatic learning of question-answer patterns based on LGs. I hope the reader will enjoy reading this journey of ideas. Key words: answer extraction, question answering, minimal logical forms, logical graphs 1