Extraction of Lexical Bundles used in Natural Language Processing Articles

Lexical bundles are indispensable for fluent academic writing. They might not constitute complete structural units but they occur very frequently in academic conversations, conference presentations and scientific articles. This paper shows how to collect a large database of lexical bundles from articles in the Natural Language Processing (NLP) domain. We first collect highly frequent N-grams from the ACL-ARC collection of NLP articles and then classify them into true or false lexical bundles using machine learning models trained from a set of manually checked bundles. In a verification experiment, our best model achieves an accuracy of 76 %. Using this model, we extract more than 18,000 lexical bundles from the ACL-ARC corpus, which we publicly release.