Proceedings of the 2003 Systems and Information Engineering Design Symposium Matthew H. Jones, Barbara E. Tawney, and K. Preston White, Jr., eds. 47 ABSTRACT What actually is done in case of text summarization in case-based reasoning terminology is that, the situation is defined as the ensemble of some consecutive sentences, and the solution is the set of the sentences selected as the outcome of the summarization process. In order to make a quality summary considering the context, a semantic un- derstanding, seems to be important. In this respect we pro- pose an approach to use a two-layered CBR approach. Re- garding this, we proposed an approach to text summarization based on two-layered case-based reasoning framework. Regarding this, the primary CBR cycle tries to make a summary of the source text, and the secondary CBR cycle tries to detect the context, and changes the bias values (fixed values) )related to the primary CBR modules. 1 INTRODUCTION Most of research in automatic text summarization has viewed in terms of extracting important concepts from the source text, building the intermediate representation and generating a coherent summary from this intermediate rep- resentation Witbrock(1999). Many methods have been proposed to extract the important concepts from a source text and to build the intermediate representation. Early methods were primarily statistical in nature and focused on word frequency to determine the most important concepts within a document Knight(2002). The opposite extreme of such statistical approaches is to attempt true “semantic understanding”, i.e. context, of the source document. Obviously, the use of deep semantic analysis offers the best opportunity to create a quality summary. The major problem with purely statistical meth- ods is that they do not account the context. To circumvent this one may offer to change the fixed parameters of the summarization approach based on the context or environmental changes. It means that one should detect the biased points of the formula/algorithm(s) and re- propose them in the way that changing the bias values be- comes easy. For example, in case of evaluating sentences for the purpose of text segmentation, one may parameterize the importance-evaluation algorithm/formula and adapts the parameters in the way it best fit to the context. For a moment, suppose the following formula which may be used in order to calculate the importance of the sentences: ( ) ( ) Query S gSim Query S Sim S Score i i i , . , . ) ( λ α + = where S i is the i-th sentence of the source text, Sim(.,.) is the statistical similarity between S i and “Query”, and gSim(.,.) is the grammatical similarity between S i and “Query”. The “Query” definition is remarkably depends on the summarization goal. Regarding the above discussion parameters α and λ should be re-valued based upon detected context, means that changing these parameters leads to different result, and should be guided based on the context. For instance when α is high, the presence frequency of words becomes more important, and the final summary contains sentences hav- ing words which are frequently used in the whole text. It is also feasible to make the other approaches like case-based reasoning, Bayesian, neural networks, or etc. sensitive to the context or environmental conditions through the above approach. Our proposed approach concentrates on applying case- based reasoning as the core methodology of text summari- zation. In this respect, we proposed an approach to text summarization based on two-layered case-based reasoning framework Badie(2002). Regarding this, the primary CBR cycle tries to make a summary of the source text, and the secondary CBR cycle tries to detect the context, and changes the bias values related to the primary CBR mod- ules. A TWO LAYERED CASE BASED REASONING APPROACH TO TEXT SUMMARIZATION, BASED ON SUMMARIZATION PATTERN Nima Reyhani Kambiz Badie Mahmoud Kharrat Info Society Dept., Iran Telecom Research Center Noth Karegar Ave. Tehran, IRAN