Citing properly in order to support concepts, claims and arguments is one of the main requirements of writing any scientific text. However, manual analysis of the input text to identify potential citation contexts is time-consuming and needs a great deal of experience. To solve this issue, in today’s citation recommendation systems, a citation worthiness identification process, also known as citation context detection, is developed as the first step in order to reduce the workload and extract an appropriate set of citation contexts. The performance of the subsequent steps is strongly dependent on the results of the citation worthiness identification task. Regarding its high importance, the current study focuses on the task of citation worthiness identification and proposes a syntax-based learning approach in order to produce a more accurate continuous representation of words. By considering both words and dependency context features, the proposed model effectively captures the functional characteristics of words and alleviates the long-distance dependency and polysemy problems. Moreover, various forms of feature representation regarding the word- and syntactic-based embeddings are presented in order to inject the best combination of them into a convolutional neural network-based classifier. The extensive experimental results on the standard dataset ACL-ARC demonstrate that the proposed model significantly improves the baselines (over 0:23 increase in F1 score) and outperforms the state-of-the-art methods (with relative improvements of over 14% in F1 score). These characteristics make the proposed model a suitable choice to be embedded in citation recommendation systems.
