| Modern, high-functionality computer applications are difficult to learn and use. Users of these applications frequently need help to complete their work. However, traditional keyword-based help systems are awkward to use and they frequently fail to satisfy their users' information needs. Allowing users to ask for help in their own words offers a more natural interaction than provided by keyword systems. Analysis of the language used in the questions facilitates the successful retrieval of answers at a far superior rate than traditional systems.; This dissertation describes SAGE, a help system developed by combining techniques gleaned from the information retrieval (IR) and artificial intelligence (AI) communities. SAGE responds to users' questions by retrieving answers from an existing collection of help text. The kernel of SAGE is based on an advanced IR technique called Latent Semantic Indexing (LSI), a variant of the vector space model. This kernel is then enhanced by adding robust natural language processing and machine learning techniques.; Both the development and evaluation of SAGE have been empirical efforts. Live user studies with a simulated system were used to gather data crucial to the design of the system. Evaluation of the system was based on live user studies and on relevance judgments for a set of naturally occurring queries.; The results of these evaluations indicate that the SAGE approach can be quite effective when deployed in realistic settings. In particular, the SAGE results have shown that given unconstrained natural language questions, answers to users' information needs can be retrieved from existing help text. |
