Automatic Term Extraction (ATE) has evolved from purely statistical or linguistic systems to hybrid approaches. More recently, machine learning is being applied to the task of identifying domain-specific vocabulary in specialised texts. Where ATE previously struggled with the data acquisition bottleneck for evaluation purposes only, the machine learning evolution means that training data are now required as well. Constructing datasets by manually annotating terms is a time- and effort-consuming task, often resulting in low inter- annotating agreement and without universally recognised guidelines. Therefore, few large datasets are available and even fewer are available in different languages and domains, annotated according to the same principles. During the first part of this project, such a dataset has been created with terms annotated in three languages (English, French and Dutch) and four domains (corruption, dressage, heart failure and wind energy), resulting in over 100k fine-grained annotations with four different labels (specific terms, common terms, out-of-domain terms and named entities).

This dataset is used to train and test a machine learning approach to term extraction. The system is built along the same principles as traditional hybrid systems: term candidates are identified by their part-of-speech patterns (patterns found in the training data) and then filtered and sorted based on other features. To this end, we combine linguistic (e.g. part-of- speech), shape (e.g. length, special characters, capitalisation) and statistical features (e.g. termhood and unithood measures). The first goal is to find out whether, using a supervised machine-learning approach, an adaptable ATE system can be trained that is able to generalise to new, unseen data in another domain, language or even differently annotated dataset. The second goal is to use this technology to learn more about the nature of terms. Which features are best at distinguishing terms from non-terms? Is it possible to train a classifier to differentiate between the different types of terms? How do term characteristics vary between different domains, languages and datasets and are there any universal, quantifiable characteristics?

During this talk, I will present these first two parts of my PhD projects: the data collection and the development of a monolingual system for ATE. The focus will be on the latter, specifically on how we intend to use the data and ATE systems to perform a qualitative analysis of terminology.