Text Mining for Deviation Detection in Financial Statemen

Abstract

This study proposes a text mining approach for detecting deviations in financial statements. Deviation detection is one of the data-mining tasks that find outliers patterns or rare data object in data. Deviation detection has wide application especially in the area of analyzing web contents, news articles and digital libraries where there is a need to retrieve anomalies that are hidden within similar documents. The rare information may produce important knowledge that led to crucial decision-making. In this study we introduce a conceptual framework of mining outliers for deviation detection in financial statements. Detecting deviations in financial statements can be beneficial to many areas such as stock forecasting, portfolio management, prediction of bankruptcy and fraud detection. Financial Statements act as a marketing and communication tool for an organization. The wealth of information in these reports is acknowledged as important for human analysts. It conveys the organizations performance to its stakeholders such as investors, creditors, auditors, financial analyst and management. Therefore, in the context of deviation detection in financial statement, any extraordinary financial reporting will be detected and highlighted as a possible new knowledge. The proposed framework includes the preprocessing and the representation of the financial statement into conceptual graphs. The preprocessing phase involves tagging the original statements into a tagged statement and parsing the tag into link grammar structure. The representation phase includes the representation of the link grammar structure into the conceptual graph. Several measures such as conceptual, relational and cumulative similarity measures identify overlap graph thus detecting the outliers in the conceptual graph. A sample of financial statement is used to generate the conceptual graphs as the training set. An example of a financial statement is used throughout this paper to describe the framework.