This thesis describes an automatic document retrieval system which analyzes the contents of documents and search requests, expressed in the natural language, and produces answers to the search request-the documents that appear to be most relevant to the request even though they may not be indexed by the exact terms of the request. A number of procedures to organize indexing systems which can provide the improvement of the total system performance are studied. Among these are functional word elimination, stem dictionary and suffix-list construction, word decomposition, thesaurus constrution, and phrase generation. Problems to minimize the dictionary storage and the problems associated with fast document search with document clustering are also examined. A practical implementation of the automatic document retrieval system is represented. The system provides four different processing methods which can be used not only to simulate an actual operating environment, but also to test the effectiveness of the various available processing methods. Retrieval performance of the system is given, demonstrating the usefulness of each method.