CoRI: Collective Relation Integration with Data Augmentation for Open Information Extraction

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4706–4716 August 1–6, 2021. ©2021 Association for Computational Linguistics 4706 CoRI: Collective Relation Integration with Data Augmentation for Open Information Extraction Zhengbao Jiang1∗, Jialong Han2, Bunyamin Sisman2, Xin Luna Dong2 Language Technologies Institute, Carnegie Mellon University1 Amazon2 zhengbaj@cs.cmu.edu {jialongh,bunyamis,lunadong}@amazon.com Abstract Integrating extracted knowledge from the Web to knowledge graphs (KGs) can facilitate tasks like question answering. We study relation in- tegration that aims to align free-text relations in subject-relation-object extractions to rela- tions in a target KG. To address the challenge that free-text relations are ambiguous, previ- ous methods exploit neighbor entities and rela- tions for additional context. However, the pre- dictions are made independently, which can be mutually inconsistent. We propose a two-stage Collective Relation Integration (CoRI) model, where the first stage independently makes can- didate predictions, and the second stage em- ploys a collective model that accesses all can- didate predictions to make globally coherent predictions. We further improve the collec- tive model with augmented data from the por- tion of the target KG that is otherwise unused. Experiment results on two datasets show that CoRI can significantly outperform the base- lines, improving AUC from .677 to .748 and from .716 to .780, respectively. 1 Introduction With its large volume, the Web has been a major resource for knowledge extraction. Open infor- mation extraction (open IE; Sekine 2006; Banko et al. 2007) is a prominent approach that harvests subject-relation-object extractions in free text with- out assuming a predefined set of relations. One way to empower downstream applications like question answering is to integrate those free-text extractions into a knowledge graph (KG), e.g., Freebase. Rela- tion integration is the first step to integrate those ex- tractions, where their free-text relations (i.e., source relations) are normalized to relations in the target KG (i.e., target relations). Only after relation in- tegration can entity linking proceed to resolve the ∗This work was performed while at Amazon. Nell Billy father Burton Marie parent parent Nell Billy father Burton Marie father father mother Source graph Predicted target graph Inference Train Unmatched KG Parallel data source relation target relation source entity target entity parent mother father mother father gender gender Barack Michelle male Malia male Malia Michelle Barack parent ↓ Figure 1: A motivating example. Trained on parallel data, a local model may suffer from sparse context for a new entity pair Nell-Marie at inference, wrongly dis- ambiguating “parent” to father instead of mother. free-text subjects and objects to their canonical en- tities in the target KG. Local Approaches. Relation integration has been studied by the natural language processing (NLP) community. With exact matching in literal form between entity names in the source graph and tar- get KG, previous methods obtain parallel data, i.e., common entity pairs, between the two graphs as in Fig. 1. Features of the entity pairs (e.g., Malia- Barack) in the source graph and their relations in the …

Content truncated to 500 words