AIFB-WebScience at SemEval-2022 Task 12: Relation Extraction First - Using Relation Extraction to Identify Entities

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1687 - 1694 July 14-15, 2022 ©2022 Association for Computational Linguistics AIFB-WebScience at SemEval-2022 Task 12: Relation Extraction First - Using Relation Extraction to Identify Entities Nicholas Popovic Karlsruhe Institute of Technology popovic@kit.edu Walter Laurito FZI Research Center for Information Technology laurito@fzi.de Michael Färber Karlsruhe Institute of Technology michael.faerber@kit.edu Abstract In this paper, we present an end-to-end joint entity and relation extraction approach based on transformer-based language models. We apply the model to the task of linking mathe- matical symbols to their descriptions in LaTeX documents. In contrast to existing approaches, which perform entity and relation extraction in sequence, our system incorporates information from relation extraction into entity extraction. This means that the system can be trained even on data sets where only a subset of all valid entity spans is annotated. We provide an ex- tensive evaluation of the proposed system and its strengths and weaknesses. Our approach, which can be scaled dynamically in computa- tional complexity at inference time, produces predictions with high precision and reaches 3rd place in the leaderboard of SemEval-2022 Task 12. For inputs in the domain of physics and math, it achieves high relation extraction macro F1 scores of 95.43% and 79.17%, respectively. The code used for training and evaluating our models is available on GitHub1. 1 Introduction Information extraction systems are a key compo- nent in making scientific literature more consum- able. With the large amount of scientific works which are constantly being published (e.g., more than 60,000 machine learning papers per year (Fär- ber, 2019)), indexing techniques that go beyond keyword searches are becoming more important. While many efforts have focused on the processing of abstracts as a way of building representations of publications (Gábor et al., 2018; Luan et al., 2018), methods processing full text documents will be needed to accurately capture their contents for use cases such as academic search and recommender systems and scientific impact quantification. The task tackled in this paper (Lai et al., 2022), consisting of linking mathematical symbols to their 1https://github.com/nicpopovic/RE1st descriptions in LaTeX documents, is a joint entity and relation extraction task. While earlier work tackled both subtasks sequentially via separate models, more recent approaches tend to use a single joint model (Luan et al., 2018; Bekoulis et al., 2018; Nguyen and Verspoor, 2019; Eberts and Ulges, 2021). In contrast to early approaches, which are based on Bi-LSTMs (Luan et al., 2018; Bekoulis et al., 2018; Nguyen and Verspoor, 2019), more recent approaches (Wadden et al., 2019; Eberts and Ulges, 2021) make use of transformer-based lan- guage models, such as BERT (Devlin et al., 2019). A key challenge in joint models is the computa- tional complexity stemming from pairwise com- parisons between entity spans required for relation extraction. Previous works tackle this using a span scoring mechanism based on a feed forward neu- ral network, which produces a score indicating the likelihood that a span is in a relation (Luan et al., 2018; Wadden …

Content truncated to 500 words