OhioState at SemEval-2018 Task 7: Exploiting Data Augmentation for Relation Classification in Scientific Papers Using Piecewise Convolutional Neural Networks

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 783–787 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics OhioState at SemEval-2018 Task 7: Exploiting Data Augmentation for Relation Classification in Scientific Papers Using Piecewise Convolutional Neural Networks Dushyanta Dhyani The Ohio State University OH, USA dhyani.2@osu.edu Abstract We describe our system for SemEval-2018 Shared Task on Semantic Relation Extraction and Classification in Scientific Papers where we focus on the Classification task. Our simple piecewise convolution neural network (PCNN) performs decently in an end to end manner. A simple inter-task data augmenta- tion significantly boosts the performance of the model. Our best-performing systems stood 8th out of 20 teams on the classification task on noisy data and 12th out of 28 teams on the classification task on clean data. 1 Introduction Relation extraction (RE) and Classification (RC) is an integral component of information extraction systems which aim to extract all the entity pairs and their relation ⟨e1, r, e2⟩ from a given text cor- pora. An alternate formulation of relation extrac- tion task focuses on identifying if a relation exists between a predefined pair of entities, and if yes classify from a given set of class relations. RE finds applications in a variety of domains, rang- ing from knowledge base construction to semantic parsing and question answering. However, the ap- plicability of existing efforts in relation extraction to scientific text calls for a quantitative and quali- tative analysis which is the aim of this shared task. 2 Related Work Existing efforts for RE range from traditional strategies (Qian et al., 2008; Bunescu and Mooney, 2006, 2005; Mintz et al., 2009; Riedel et al., 2010) to more recent end to end deep learn- ing based methods (Zeng et al., 2014, 2015; Lin et al., 2016; Wu et al., 2017) that are more suit- able in situations where a lot of training data is available. While a majority of efforts in the RE community are specifically focused towards using distantly supervised data and reduce the associ- ated noise, their discussion is not relevant to the current scenario. The most relevant work is that of (Zeng et al., 2014) who demonstrated the effi- cacy of convolution neural networks for relation classification and (Zeng et al., 2015) who further enhanced the architecture by proposing the piece- wise max-pooling strategy. 3 Task Description The semantic relation extraction and classification in scientific papers task (G´abor et al., 2018) aims at identifying semantic relations expressed by en- tity pairs in scientific literature. The contest is fur- ther divided into three subtasks, where the first two focus on classification of varying nature of data and the third focuses on extraction task. Since our submitted systems focused only on the classifica- tion task, we would from here on discuss mostly about the classification sub-tasks. 3.1 Dataset The data contains titles and abstracts of papers from ACL Anthology Corpus where entity men- tions are either manually annotated (Subtask 1.1 and Subtask 2) or heuristically (Subtask 1.2) …

Content truncated to 500 words