Multilingual Entity and Relation Extraction Dataset and Model

Abstract

We present a novel dataset and model for a multilingual setting to approach the task of Joint Entity and Relation Extraction. The SMi-LER dataset consists of 1.1 M annotated sentences, representing 36 relations, and 14 languages. To the best of our knowledge, this is currently both the largest and the most comprehensive dataset of this type. We introduce HERBERTa, a pipeline that combines two independent BERT models: one for sequence classification, and the other for entity tagging. The model achieves micro F 1 81.49 for English on this dataset, which is close to the current SOTA on CoNLL, SpERT.

DiS-ReX: A Multilingual Dataset for Distantly Supervised Relation Extraction

“…Contemporary to our work, other multilingual RE datasets and methods are being developed. These include a dataset for joint entity and relation extraction (Seganti et al, 2021), a model for multilingual KB completion (Singh et al, 2021), and an approach for automatic construction of crosslingual training data for Open IE (Kolluru et al, 2022). Our proposed dataset, DiS-ReX, has already been used for further research on the Multilingual DS-RE task (Rathore et al, 2022).…”

What do You Mean by Relation Extraction? A Survey on Datasets and Study on Scientific Relation Classification

“…The multi-lingual dimension is gaining more interest for RE. Following this trend, Seganti et al (2021) presented SMiLER, a multilingual dataset (14 languages) from Wikipedia with relations belonging to nine domains.…”

“…The last step (RC) is usually a multi-class classification to assign a relation type r to the positive samples from the previous step. Some studies merge RI and RC (Seganti et al, 2021) into one step, by adding a no-relation (no-rel) label. Other studies instead reduce the task to RC, and assume there exists a relation between two entities and the task is to determine the type (without a no-rel label).…”

“…2019). Similarly, since the advent of contextualized representations (Peters et al, 2018;Devlin et al, 2019), BERT-like representations are commonly used (Seganti et al, 2021), but non-contextualized embeddings (i.e., GloVe, fastText) are still widely adopted (Yao et al, 2019;Huang et al, 2021b). We compare the best CNN setup to fine-tuning a full transformer model.…”