Refining Readability by Tackling Redundancy and Time through Modified Rule based Extractive Summarization Approach for Low Resource Indian Languages-Hindi andPunjabi

Abstract

Abstract A rule based extractive text summarization system is proposed to obtain comprehensible summarized text of input Hindi and Punjabi text documents. The input document to be summarized can be any handwritten text or any text from internet saved in .rtf format. It works on Hindi and Punjabi documents and summarizes text of any length i.e. system summarizes text irrespective of size. The extensive rule based approach is applied consisting preprocessing and processing phases. Each phase consists of rules which are to be followed to get the desired concise summary. In the preprocessing phase, rules of Tokenization, Stop word removal and segmentation are applied whereas in processing phase rules such as Monetary values, Measurement Values, Special Symbols, Equations, Figure Number in sentences are applied. Also, other features such as frequently occurring words and maximum length word are added to make the system output comprehendible. The system is made user friendly by proposing a graphical user interface (GUI). The generated summary by the proposed system is compared with the summary of documents generated by human experts and is evaluated in terms of parameters such as Compression Ratio and Execution Time. Further, accuracy is calculated based on Precision, Recall and F-score. It has been observed that the compression ratio obtained lies in the range of 39–50% for 200, 300 and 400 words paragraphs. The execution time is quite less which do varies with the size of the paragraph entered as the summary size varies. In general, it has been observed that the execution time for 200 words paragraph takes 1519.147 milliseconds, for 300 words paragraph the time is 3022.825 milliseconds whereas for 400 words size paragraph, the time taken is 4023.123 milliseconds. The time is quite less as compared to human generated summary. The accuracy obtained in each case is above 95% which is high as compared to humans and other state of art methods. This shows that the system is robust enough and provides definitive results.