Revisiting Document-Level Relation Extraction with Context-Guided Link Prediction
Authors: Monika Jain, Raghava Mutharaju, Ramakanth Kavuluru, Kuldeep Singh
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach on three benchmark datasets: Doc RED, Re Doc RED, and DWIE. The results indicate that our proposed method outperforms the state-of-the-art models and suggests that incorporating context-based link prediction techniques can enhance the performance of document-level relation extraction models. |
| Researcher Affiliation | Collaboration | Monika Jain1, Raghava Mutharaju1, Ramakanth Kavuluru2, Kuldeep Singh3 1Indraprastha Institute of Information Technology, Delhi, India 2University of Kentucky, Lexington, Kentucky, United States 3Cerence Gmb H and Zerotha Research, Germany |
| Pseudocode | No | The paper describes the approach and methods using text and mathematical equations but does not provide pseudocode or a clearly labeled algorithm block. |
| Open Source Code | Yes | The code and the models are available at https://github.com/kracr/document-levelrelation-extraction. |
| Open Datasets | Yes | The proposed model is evaluated on three widely-used public datasets 1) Doc RED (Yao et al. 2019), containing 50,503 triples, 5053 documents and 96 relation 2) Re Doc RED (Ma, Wang, and Okazaki 2023) containing 120,664 triples, 5053 documents, 96 relation and 3) DWIE (Zaporojets et al. 2021). |
| Dataset Splits | Yes | Ign F1 is measured by removing relations in the annotated training set from the development and test sets. The mean and standard deviation of F1 and Ign F1 on the dev set are reported for three training runs. We report the official test score for Doc RED on the best checkpoint on the dev set. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper mentions software components like GloVe, Bidirectional LSTM, BERT-Based model, R-GCN, AdamW, Adam, and Dist Mult, but does not provide specific version numbers for any of them. |
| Experiment Setup | Yes | We employ uncased BERT-Based model (768d) as an encoder with a learning rate 1e-3. We used Adam W as an optimizer, and the learning rate is 1e 3. R-GCN is used as an encoder with a single encoding layer (200d) embeddings for the link prediction model. We regularize the encoder through edge dropout applied before normalization, with a dropout rate of 0.2 for self-loops and 0.4 for other edges. We apply l2 regularization to the decoder with a penalty of 0.01. Adam (Kingma and Ba 2015) is used as an optimizer, and the model is trained with 100 epochs using a learning rate of 0.01. ... The N-hop path length of the context varies from 1 to 4. ... We have used beam size of 128 for beam search for all three datasets. |