Multilevel Entity-Informed Business Relation Extraction

Abstract

This paper describes a business relation extraction system that combines contextualized language models with multiple levels of entity knowledge. Our contributions are three-folds: (1) a novel characterization of business relations, (2) the first large English dataset of more than 10k relation instances manually annotated according to this characterization, and (3) multiple neural architectures based on BERT, newly augmented with three complementary levels of knowledge about entities: generalization over entity type, pre-trained entity embeddings learned from two external knowledge graphs, and an entity-knowledgeaware attention mechanism. Our results show an improvement over many strong knowledge-agnostic and knowledge-enhanced state of the art models for relation extraction.

How Can a Teacher Make Learning From Sparse Data Softer? Application to Business Relation Extraction

“…company-customer, company-partner) at the sentence level (Zhao et al, 2010). For example, from the sentence in (1), extracted from BIZREL dataset (Khaldi et al, 2021), a relation extraction system can infer that the company Airbus is a supplier for the company Inmarsat. Recent works for BRE rely on supervised approaches, where neural models are trained on annotated datasets for business relations (Collovini et al, 2020;De Los Reyes et al, 2021;Khaldi et al, 2021).…”

“…For example, from the sentence in (1), extracted from BIZREL dataset (Khaldi et al, 2021), a relation extraction system can infer that the company Airbus is a supplier for the company Inmarsat. Recent works for BRE rely on supervised approaches, where neural models are trained on annotated datasets for business relations (Collovini et al, 2020;De Los Reyes et al, 2021;Khaldi et al, 2021). In general, supervised approaches consider relation extraction (RE) as a multi-class classification problem where each class corresponds to a predefined relation type (Zhang et al, 2017;Wu, 2019).…”

“…In general, supervised approaches consider relation extraction (RE) as a multi-class classification problem where each class corresponds to a predefined relation type (Zhang et al, 2017;Wu, 2019). In addition to the set of positive relations (henceforth PR) which corresponds to the taxonomy of relations of interest (like hypernymy, meronymy, and cause-effect relationships), most popular datasets manually annotated either for generic (e.g., SemEval-2010 Task 8 (Hendrickx et al, 2010), TACRED (Zhang et al, 2017)) or domain specific relations (e.g., ChemProt (Krallinger et al, 2017), BizRel (Khaldi et al, 2021)) include a negative relation (henceforth NR) to account either for the absence of a relation between two target entities (see NO-RELATION in TACRED), or any other types of relations not present in the annotation scheme (see OTHERS in SemEval-2010 and BizRel). NRs share two main characteristics: (C1) they have irregular and unstable linguistic realizations and (C2) are often over-represented making PR hard to predict due to the highly imbalanced nature of the problem (see the ratio of NR in Table 1).…”

“…(3) Multitask learning where auxiliary tasks help the main task to improve performances of under-represented classes (Khaldi et al, 2021;Wang and Hu, 2020).…”

“…Indeed, NR can have unstable patterns, and can share similar linguistic realizations with PR. For example, the SemEval 2010 Task 8 dataset (Hendrickx et al, 2010) includes the OTHERS relations for near misses of PR as NR instances, while in the BIZREL dataset (Khaldi et al, 2021) (Song et al, 2021). In this paper, we aim to continue these efforts by proposing a new knowledge distillation approach based on binary soft labels supervision (BSLS).…”