Resolving Citation Links With Neural Networks

This work demonstrates how neural network models (NNs) can be exploited toward resolving citation links in the scientific literature, which involves locating passages in the source paper the author had intended when citing the paper. We look at two kinds of models: triplet and binary. The triplet network model works by ranking potential candidates, using what is generally known as the triplet loss, while the binary model tackles the issue by turning it into a binary decision problem, i.e., by labeling a candidate as true or false, depending on how likely a target it is. Experiments are conducted using three datasets developed by the CL-SciSumm project from a large repository of scientific papers in the Association for Computational Linguistics (ACL) repository. The results find that NNs are extremely susceptible to how the input is represented: they perform better on inputs expressed in binary format than on those encoded using the TFIDF metric or neural embeddings of specific kinds. Furthermore, in response to a difficulty NNs and baselines faced in predicting the exact location of a target, we introduce the idea of approximately correct targets (ACTs) where the goal is to find a region which likely contains a true target rather than its exact location. We show that with the ACTs, NNs consistently outperform Ranking SVM and TFIDF on the aforementioned datasets