Domain-specific hierarchical text classification for supporting automated environmental compliance checking

Automated environmental compliance checking requires automated extraction of rules from environmental regulatory textual documents such as energy conservation codes and EPA regulations. Automated rule extraction requires complex text processing and analysis for information extraction and subsequent formalization of the extracted information into computer-processable rules. In the proposed automated compliance checking (ACC) approach, the text is first classified into predefined categories before information extraction (IE). The advantages are that irrelevant text will be filtered out during text classification (TC) and text with similar semantic meaning will be grouped, thereby improving the efficiency and accuracy of further IE and compliance reasoning (CR). The categories used for TC are predefined in a semantic TC topic hierarchy, and the classified text is subsequently used in semantic IE and semantic CR. This paper presents the proposed machine learning (ML)-based TC algorithm for classifying clauses in environmental regulatory documents based on the TC topic hierarchy. In developing the algorithm, different text preprocessing techniques, ML algorithms, and performance improvement strategies were tested and used. The final TC algorithm was tested on 10 environmental regulatory documents and evaluated in terms of precision and recall. The algorithm achieved approximately 97 and 84% average recall and precision, respectively, on the testing data.