ANALYSIS OF ANNUAL REPORTS BY ADVANCED NEURAL NETWORK METHODS
Professor: Ari Visa
The goal of the project was to test if it is possible to find any correlation between the text part and the corresponding economical figures of annual reports. The problem is complicated because depending on the report there is more or less correlation. There might also be some time delays between the proclamation and the actual figures. The team was created by experts in the economics, in the management, and in the neurocomputing. The problem was studied by four hypothesises. As a reference time series of the economical figures of the actual companies were considered. The analysed statements of the annual reports were compared with the behaviour of the time series. As an answer to the text interpretation problem a new technology based on multilevel hierarchies of Self-Organizing Feature Maps and on smart encoding of words was proposed. The results are interesting and promising. They are supporting the hypothesises but one should be careful with the final conclusions. However, in our second test case, where we analysed fault reports in electrical power industry, the results were interesting in such a way that the work will continue at least in that field.
As an answer to the text interpretation problem a new technology was proposed. The Technology is based on multilevel hierarchies of Self-Organizing Feature Maps and on smart encoding of words.The encoding of the word is language independent. The levels of the hierarchy are word, sentence, and paragraph maps. Our experiments with text documents (annual reports) show that with text documents it is possible to achieve similar results as with analysis of economical figures. Besides that it is possible both to find similar documents and to separate between different types of documents. This is also true considering paragraphs within a document. The technology might find some applications in the field of intellectual property rights concerning text documents. In our second test case, where we analysed fault reports in electrical power industry, the results were promising. It seems that the work will continue with more specific applications.
The research work was done in Department of Information Science at Lappeenranta University of Technology and in Laboratory of Information Systems at Åbo Akademi University, Finland during 1.March 1998 - 31. December 1999. The budget of this two years project was roughly 1.2 MFIM, and the amount of work was about 44 person months. The work was carried out in close co-operation with the following industrial companies:
The project manager was professor Ari Visa.
The publishing of the results is just going on. To date four papers have been published and four papers have been submitted. The information about the future publications can be found from this web page.
Refereed International Conference Papers