Semantic Text Segment Classification of Structured Technical Content

Abstract

Semantic tagging in technical documentation is an important but error-prone process, with the objective to produce highly structured content for automated processing and standardized information delivery. Benefits thereof are consistent and didactically optimized documents, supported by professional and automatic styling for multiple target media. Using machine learning to automate the validation of the tagging process is a novel approach, for which a new, high-quality dataset is provided in ready-to-use training, validation and test sets. In a series of experiments, we classified ten different semantic text segment types using both traditional and deep learning models. The experiments show partial success, with a high accuracy but relatively low macro-average performance. This can be attributed to a mix of a strong class imbalance, and high semantic and linguistic similarity among certain text types. By creating a set of context features, the model performances increased significantly. Although the data was collected to serve a specific use case, further valuable research can be performed in the areas of document engineering, class imbalance reduction, and semantic text classification.