Fine-grained change detection in structured text documents

Abstract

Detecting and understanding changes between document revisions is an important task. The acquired knowledge can be used to classify the nature of a new document revision or to support a human editor in the review process. While purely textual change detection algorithms offer fine-grained results, they do not understand the syntactic meaning of a change. By representing structured text documents as XML documents we can apply tree-to-tree correction algorithms to identify the syntactic nature of a change. Many algorithms for change detection in XML documents have been propsed but most of them focus on the intricacies of generic XML data and emphasize speed over the quality of the result. Structured text requires a change detection algorithm to pay close attention to the content in text nodes, however, recent algorithms treat text nodes as black boxes. We present an algorithm that combines the advantages of the purely textual approach with the advantages of tree-to-tree change detection by redistributing text from non-overlapping common substrings to the nodes of the trees. This allows us to not only spot changes in the structure but also in the text itself, thus achieving higher quality and a fine-grained result in linear time on average. The algorithm is evaluated by applying it to the corpus of structured text documents that can be found in the English Wikipedia.