Composing Structure-Aware Batches for Pairwise Sentence Classification

Findings of the Association for Computational Linguistics: ACL 2022, pages 257 - 268 May 22-27, 2022 c⃝2022 Association for Computational Linguistics EIDER: Empowering Document-level Relation Extraction with Efficient Evidence Extraction and Inference-stage Fusion Yiqing Xie, Jiaming Shen, Sha Li, Yuning Mao, Jiawei Han University of Illinois at Urbana-Champaign, IL, USA {xyiqing2, js2, shal2, yuningm2, hanj}@illinois.edu Abstract Document-level relation extraction (DocRE) aims to extract semantic relations among entity pairs in a document. Typical DocRE methods blindly take the full document as input, while a subset of the sentences in the document, noted as the evidence, are often sufficient for humans to predict the relation of an entity pair. In this paper, we propose an evidence- enhanced framework, EIDER, that empowers DocRE by efficiently extracting evidence and effectively fusing the extracted evidence in in- ference.1 We first jointly train an RE model with a lightweight evidence extraction model, which is efficient in both memory and runtime. Empirically, even training the evidence model on silver labels constructed by our heuristic rules can lead to better RE performance. We further design a simple yet effective inference process that makes RE predictions on both ex- tracted evidence and the full document, then fuses the predictions through a blending layer. This allows EIDER to focus on important sen- tences while still having access to the complete information in the document. Extensive exper- iments show that EIDER outperforms state-of- the-art methods on three benchmark datasets (e.g., by 1.37/1.26 Ign F1/F1 on DocRED). 1 Introduction Relation extraction (RE) is the task of extracting se- mantic relations among entities within a given text, which has abundant applications such as knowl- edge graph construction, question answering, and biomedical text analysis (Yu et al., 2017; Shi et al., 2019; Trisedya et al., 2019). Prior studies mostly focus on predicting the relation between two entity mentions in a single sentence. However, in reality, an entity may have multiple mentions throughout a document. It is also common that a relation can only be inferred given multiple sentences as the 1Our code is available at https://github.com/ Veronicium/Eider Head:Hero of the Day Tail:the United States Rel:[country of origin] GT evidence sentences: [1,10] Extracted evidence: [1,10] Original document as input: [1] Load is the sixth studio album by the American heavy metal band Metallica, released on June 4, 1996 by Elektra Records in the United States … [9] It was certified 5×platinum … for shipping five million copies in the United States. [10] Four singles—"Hero of the Day", "Until It Sleeps", "Mama Said", and "King Nothing" — were released as part of the marketing campaign for the album. Prediction scores: NA: 17.63 country of origin: 14.79 Extracted evidence as input: [1] Load is the sixth studio album … released … in the United States … [10] Four singles — "Hero of the Day", … were released … for the album. Prediction scores: country of origin: 18.31 NA: 13.45 Final prediction of our model: country of origin (✓) Figure 1: A test sample in the DocRED dataset …

Content truncated to 500 words