[Paper Review] Recurrent Interaction Network for Jointly Extracting Entities and Classifying Relations

Kai Sun, Richong Zhang+3 more|arXiv (Cornell University)|May 1, 2020

Topic Modeling32 references5 citations

TL;DR

This paper proposes a Recurrent Interaction Network (RIN) for joint entity recognition and relation classification, dynamically modeling task-specific interactions through a recurrent multi-layer architecture to improve feature representation. The model achieves state-of-the-art F1 scores of 84.7 on NYT and 77.0 on WebNLG, outperforming existing multi-task learning methods by explicitly learning cross-task dependencies.

ABSTRACT

The idea of using multi-task learning approaches to address the joint extraction of entity and relation is motivated by the relatedness between the entity recognition task and the relation classification task. Existing methods using multi-task learning techniques to address the problem learn interactions among the two tasks through a shared network, where the shared information is passed into the task-specific networks for prediction. However, such an approach hinders the model from learning explicit interactions between the two tasks to improve the performance on the individual tasks. As a solution, we design a multi-task learning model which we refer to as recurrent interaction network which allows the learning of interactions dynamically, to effectively model task-specific features for classification. Empirical studies on two real-world datasets confirm the superiority of the proposed model.

Motivation & Objective

• To address the limitation of conventional multi-task learning models that rely on static shared representations and fail to model explicit interactions between entity recognition and relation classification.
• To improve classification performance on both entity recognition and relation classification by enabling dynamic, iterative information exchange between the two tasks.
• To overcome error propagation in pipeline methods and the restrictive flat structure of existing MTL models through a graph-structured, recurrent architecture.
• To empirically validate that explicit, dynamic interaction between tasks leads to better feature learning and improved performance on benchmark datasets.

Proposed method

• The RIN model uses a recurrent multi-layer structure with alternating updates between shared features and task-specific features, enabling iterative refinement of representations.
• At each interaction layer, shared features are updated using both task-specific features from entity recognition and relation classification, allowing bidirectional information flow.
• The model incorporates POS tags and pre-trained GloVe embeddings as input features, with residual connections to stabilize training.
• Task-specific features are refined through gated mechanisms that selectively update based on cross-task feedback.
• The architecture is designed as a graph-structured MTL framework, where information flows dynamically between the ER and RC heads across multiple layers.
• The model is trained end-to-end using cross-entropy loss for both tasks, with joint optimization to enhance feature discrimination.

Experimental results

Research questions

• RQ1 Can dynamic, iterative interaction between entity recognition and relation classification tasks improve performance beyond static shared representations?
• RQ2 Does explicitly modeling cross-task dependencies lead to better feature learning and classification accuracy?
• RQ3 How does the proposed RIN model compare to conventional MTL and pipeline approaches in terms of F1 score and robustness on overlapping relational facts?
• RQ4 What is the contribution of each component (e.g., interaction mechanism, POS features) to the overall performance?

Key findings

• The RIN model achieves a new state-of-the-art F1 score of 84.7 on the NYT dataset and 77.0 on the WebNLG dataset under exact match evaluation.
• Ablation studies show that removing the interaction mechanism (RIN w/o interaction) causes the largest performance drop, confirming its critical role in capturing complex task correlations.
• The model performs better when the entity recognition module can access relation classification feedback (RIN w/o ER), indicating that RC provides valuable context for ER.
• The ablation without POS tags (RIN w/o POS) shows only a minor drop in performance, suggesting that POS features are not essential for the model’s success.
• Case studies demonstrate that RIN correctly identifies overlapping relational facts and corrects entity mispredictions by leveraging cross-task feedback, unlike models without interaction.
• The RIN model effectively learns discriminating features through iterative refinement, which is essential for high-accuracy classification in joint extraction tasks.