DocRED: A Large-Scale Document-Level Relation Extraction Dataset

Abstract

Multiple entities in a document generally exhibit complex inter-sentence relations, and cannot be well handled by existing relation extraction (RE) methods that typically focus on extracting intra-sentence relations for single entity pairs. In order to accelerate the research on document-level RE, we introduce DocRED, a new dataset constructed from Wikipedia and Wikidata with three features: (1) DocRED annotates both named entities and relations, and is the largest humanannotated dataset for document-level RE from plain text; (2) DocRED requires reading multiple sentences in a document to extract entities and infer their relations by synthesizing all information of the document; (3) along with the human-annotated data, we also offer large-scale distantly supervised data, which enables DocRED to be adopted for both supervised and weakly supervised scenarios. In order to verify the challenges of documentlevel RE, we implement recent state-of-the-art methods for RE and conduct a thorough evaluation of these methods on DocRED. Empirical results show that DocRED is challenging for existing RE methods, which indicates that document-level RE remains an open problem and requires further efforts. Based on the detailed analysis on the experiments, we discuss multiple promising directions for future research. We make DocRED and the code for our baselines publicly available at https: //github.com/thunlp/DocRED. * indicates equal contribution

HIN: Hierarchical Inference Network for Document-Level Relation Extraction

“…For document-level RE, the input is a document with annotated entities, as well as multiple occurrences of each entity, i.e., entity mentions, the goal is to identify all the related entity pairs in the document. Following [15], we transform RE into a classification problem. We use upper case letters to represent entities (E 1 , · · · , E m ) and lower case letters to represent mentions (e 1 , · · · , e m ).…”

“…To evaluate the effectiveness of our model, we use the DocRED dataset [15], which is the largest human-annotated document-level RE dataset constructed from Wikidata and Wikipedia. DocRED contains over 5,053 documents, 40,276 sentences, 132,375 entities and 96 frequent relation types.…”

“…We compare our model against the following document-level RE baselines: CNN/LSTM/BiLSTM-RE: They first encode a document into a hidden state vector sequence with CNN/LSTM/BiLSTM as encoder, and then predict relations for each entity pair by feeding them into a bilinear function [15]. Context-Aware: It uses an LSTM-based encoder to jointly learn representations for all relations in the context, and then combines other context relations with target relation to make the final prediction [12].…”

“…Recently, there has been increasing interest in document-level RE. Yao et al [15] proposed a large-scale human-annotated document-level RE dataset, DocRED, and first compute the representations for all entities then predict relations for each entity pair by feeding them into a bilinear function. Wang et al [13] used BERT to encode the document, it also used bilinear layer to predict the relation between entity pairs, but it modelled the document-level RE through a two-step process.…”

“…Obviously, most traditional sentence-level RE models often fail to generalize extraction to this situation. To move RE forward from sentence level to document level, many efforts have been made [13,15], but most previous methods used only entity-level information and this is not adequate. Thus, there are still some deep-seated problems unsolved in document-level RE.…”