Formalising Document Structure and Automatically Recognising Document Elements: A Case Study on Automobile Repair Manuals

Abstract

Correct understanding of document structures is vital to enhancing various practices and applications, such as authoring and information retrieval. In this paper, with a focus on procedural documents of automobile repair manuals, we report a formalisation of document structure and an experiment to automatically classify sentences according to document structure. To formalise document structure, we first investigated what types of content are included in the target documents. Through manual annotation, we identified 26 indicative content categories (content elements). We then employed an existing standard for technical documents—the Darwin Information Typing Architecture—and formalised a detailed document structure for automobile repair tasks that specifies the arrangement of the content elements. We also examined the feasibility of automatic content element recognition in given documents by implementing sentence classifiers using multiple machine learning methods. The evaluation results revealed that support vector machine-based classifiers generally demonstrated high performance for in-domain data, while convolutional neural network-based classifiers are advantageous for out-of-domain data. We also conducted in-depth analyses to identify classification difficulties.