Guide the Many-to-One Assignment: Open Information Extraction via IoU-aware Optimal Transport

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4971–4984 July 9-14, 2023 ©2023 Association for Computational Linguistics Guide the Many-to-One Assignment: Open Information Extraction via IoU-aware Optimal Transport Kaiwen Wei1,2∗, Yiran Yang1,2∗, Li Jin1,2†, Xian Sun1,2, Zequn Zhang1,2, Jingyuan Zhang3, Xiaoyu Li1,2, Linhao Zhang1,2, Jintao Liu1,2, Zhi Guo1,2 1Key Laboratory of Network Information System Technology, Aerospace Information Research Institute, Chinese Academy of Sciences 2School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences 3DAMO Academy, Alibaba Group {weikaiwen19,yangyiran19}@mails.ucas.ac.cn, jinlimails@gmail.com Abstract Open Information Extraction (OIE) seeks to extract structured information from raw text without the limitations of close ontology. Re- cently, the detection-based OIE methods have received great attention from the community due to their parallelism. An essential step of those models is how to assign ground truth la- bels to the parallelly generated tuple proposals, which remains under-exploited. The commonly utilized Hungarian algorithm for this procedure is restricted to handling one-to-one assignment among the desired tuples and tuple propos- als, which ignores the correlation between proposals and affects the recall of the models. To solve this problem, we propose a dynamic many-to-one label assignment strategy named IOT. Concretely, the label assignment process in OIE is formulated as an Optimal Transport (OT) problem. We leverage the intersection- over-union (IoU) as the assignment quality measurement, and convert the problem of finding the best assignment solution to the one of solving the optimal transport plan by maximizing the IoU values. To further utilize the knowledge from the assignment, we design an Assignment-guided Multi-granularity (AM) loss by simultaneously considering word-level and tuple-level information. Experiment results show the proposed method outperforms the state-of-the-art models on three benchmarks. 1 Introduction Open Information Extraction (OIE) aims to extract structured information from the given text without the restriction of pre-defined ontology schema, and it is typically formed as a tuple ⟨subject, relation, object⟩ (Yates et al., 2007). For example, given the sentence "Dr. Pim played for Ireland against England", an OIE system needs to extract ⟨Dr. Pim, *Equal contribution. †Corresponding author. (b) Hungarian Algorithm Assignment Matrix Optimal Transport pr 1 pr 2 pr 3 pr 4 pr 5 gt 1 0.1 0.53 0.52 0.3 0.1 gt 2 0.8 0.4 0.2 0.2 0.1 pr 1 pr 2 pr 3 pr 4 pr 5 gt 1 0.09 0.81 0.8 0.6 0.16 gt 2 0.9 0.14 0.1 0.3 0.14 bg 0.01 0.05 0.1 0.1 0.7 (a) IoU Matrix pr 1 pr 2 pr 3 pr 4 pr 5 gt 1 0 1 0 0 0 gt 2 1 0 0 0 0 Hungarian Algorithm (c) Optimal Transport Solution Matrix (one pr to one gt) (many prs to one gt) Input Sentence: Dr. Pim played for Ireland against England Tuple Predictions: (Pim, played against, England) (Dr. Pim, played, Ireland) (Dr., played for, Ireland) ... Ground Truth Labels (gts): (Dr. Pim, played for, Ireland) (Dr. Pim, played against, England) Figure 1: The toy example of the label assignment pro- cess from …

Content truncated to 500 words