Multilingual Entity and Relation Extraction Dataset and Model

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 1946–1955 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 1946 Multilingual Entity and Relation Extraction Dataset and Model Alessandro Seganti∗2, Klaudia Firl ˛ag1, Helena Skowro´nska*3, Michał Satława1, and Piotr Andruszkiewicz1,4 1Samsung R&D Institute Poland 2Equinix 3NextSell, ODC Group 4Warsaw University of Technology alessandro.seganti@gmail.com {k.firlag, m.satlawa, p.andruszki2}@samsung.com arvala@wp.pl, p.andruszkiewicz@ii.pw.edu.pl Abstract We present a novel dataset and model for a multilingual setting to approach the task of Jo- int Entity and Relation Extraction. The SMi- LER dataset consists of 1.1 M annotated sen- tences, representing 36 relations, and 14 lan- guages. To the best of our knowledge, this is currently both the largest and the most com- prehensive dataset of this type. We introduce HERBERTa, a pipeline that combines two in- dependent BERT models: one for sequence classification, and the other for entity tagging. The model achieves micro F1 81.49 for En- glish on this dataset, which is close to the cur- rent SOTA on CoNLL, SpERT. 1 Introduction The majority of the – constantly growing – amo- unt of openly accessible knowledge is locked in unstructured text, and hence inefficiently utilized by any systems. The NLP tasks related to this problem include Information Extraction, Relation Extraction, Named Entity Retrieval, as well as Joint Entity and Relation Extraction. Our contribution is twofold. First, we present SMiLER (Samsung MultiLingual Entity and Rela- tion Extraction dataset): an open-domain corpus of annotated sentences, created for the Joint Entity and Relation Extraction task. With 1.1 M senten- ces, 36 relation types, and 14 languages, SMiLER seems to be both the largest and the most diversified corpus for the task in existence, to the best of our knowledge. The corpus was semi-automatically created from Wikipedia and DBpedia, and partly checked by linguists. Our second contribution is HERBERTa – Hybrid Entity and Relation extraction BERT for a multi- ∗Work done while at Samsung R&D Institute Poland. lingual setting that consists of two independently pretrained BERT models (Devlin et al., 2018). The first one classifies the input sequence as belonging to one of our 36 relations (including no_relation). Its output – the relation – is then fed to the second BERT, together with the same input sequence. The second model performs entity tagging and outputs two spans for the two entities selected from the in- put sequence. Our model is close to SpERT (Eberts and Ulges, 2019), the current state-of-the-art for Jo- int Entity and Relation Extraction, achieving micro F1 81.49 for English on the presented dataset. The SMiLER corpus and the source code is ava- ilable at https://github.com/samsungnlp/smiler/. 2 Related Work Available Datasets The datasets that are com- monly used for the task of Joint Entity and Relation Extraction are still insufficient in size and diver- sification; furthermore, they are all monolingual English corpora. For instance, CONLL04 (Roth and Yih, 2004) has only 1.7k annotated sentences and 5 relations. ADE (Gurulingappa et al., 2012) is larger, with almost 21k sentences, …

Content truncated to 500 words