Semi-supervised Relation Extraction via Data Augmentation and Consistency-training

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1112–1124 May 2-6, 2023 ©2023 Association for Computational Linguistics Semi-supervised Relation Extraction via Data Augmentation and Consistency-training Komal K. Teru The Vanguard Group komal_teru@vanguard.com Abstract Due to the semantic complexity of the Relation extraction (RE) task, obtaining high-quality hu- man labelled data is an expensive and noisy pro- cess. To improve the sample efficiency of the models, semi-supervised learning (SSL) meth- ods aim to leverage unlabelled data in addition to learning from limited labelled data points. Recently, strong data augmentation combined with consistency-based semi-supervised learn- ing methods have advanced the state of the art in several SSL tasks. However, adapting these methods to the RE task has been chal- lenging due to the difficulty of data augmen- tation for RE. In this work, we leverage the recent advances in controlled text generation to perform high quality data augmentation for the RE task. We further introduce small but significant changes to model architecture that allows for generation of more training data by interpolating different data points in their latent space. These data augmentations along with consistency training result in very competitive results for semi-supervised relation extraction on four benchmark datasets. 1 Introduction Relation extraction is one of the essential compo- nents in constructing structured knowledge bases (Luan et al., 2018), performing interpretable ques- tion answering (Sun et al., 2021), improving web search, and many other information extraction pipelines. It aims to discover the semantic relation between a given head entity and tail entity based on the context in the input sentence. For example, given a sentence "The battle led to panic on the frontier, and settlers in the surrounding counties fled.", the goal is to extract the Cause-Effect relation between the head entity ‘battle’ and the tail entity ‘panic’. The RE task requires a high level of language understanding and involves a sig- nificant level of semantic complexity (Bach and Badaskar, 2007). Due to this semantic complexity Input sentence: The battle led to panic on the frontier, and settlers in the surrounding counties fled. Synonym-replacement: The struggle cause scare on the frontier, and settlers in the surrounding counties fly. LM-based augmentation: The battle reduced panic on the frontier, and settlers in the surrounding counties relaxed. Vanilla BT: The war caused panic at the border, and residents of the nearby counties fled. Our constrained BT: The battle sparked panic at the border, with residents fleeing in surrounding counties. Figure 1: Different data augmentation techniques ap- plied to a sample datapoint from SemEval dataset. Ex- isting methods replace the head/tail entities (highlited in red), change the original meaning or do not give very fluent paraphrases. it often requires extensive and highly skilled human involvement to obtain good quality labelled data, making data collection an expensive and noisy pro- cess. Unsurprisingly, because of the same semantic complexity of the task, models typically require large amounts of labelled data to give production- ready performance. A common strategy to improve …

Content truncated to 500 words