Tagging terms in text

Abstract As with many tasks in natural language processing, automatic term extraction (ATE) is increasingly approached as a machine learning problem. So far, most approaches to ATE broadly follow the traditional hybrid methodology, by first extracting list of unique candidate terms, and classifying these candidates based on predicted probability that they are valid terms. However, rise neural networks word embeddings, next development might be towards sequential approaches, i.e., each occurrence token within its original context. To test validity such for ATE, two methodologies were developed, evaluated, compared: one feature-based conditional random fields classifier embedding-based recurrent network. An additional comparison was added interpretation approach. All systems trained evaluated identical data multiple languages domains identify their respective strengths weaknesses. The proven network even outperformed more Interestingly, combination can outperform all them separately, showing new ways push state-of-the-art ATE.