Question answering using sentence parsing and semantic network matching

The paper describes a question answering system for German called InSicht. All documents in the system are analyzed by a syntactico-semantic parser in order to represent each document sentence by a semantic network. A question sent to InSicht is parsed yielding its semantic network representation and its sentence type. The semantic network is expanded by applying equivalence rules, implicational rules, and concept variations based on semantic relations in computer lexicons and other knowledge sources. During the search stage, every semantic network generated for the question is matched with semantic networks for document sentences. If a match succeeds, an answer is generated from the matching semantic network for the supporting document. InSicht is evaluated on the QA@CLEF 2004 test set. A hierarchy of problem classes is proposed and a sample of suboptimally answered questions is annotated with these problem classes. Finally, some conclusions are drawn, main problems are identified, and directions for future work as suggested by these problems are indicated.