Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improved Relation Classification by Deep Recurrent Neural Networks with Data Augmentation

Yan Xu, Ran Jia, Lili Mou, Ge Li, Yunchuan Chen, Yangyang Lu, Zhi Jin

TL;DR

This work addresses relation classification by introducing Deep Recurrent Neural Networks (DRNNs) built on the Shortest Dependency Path (SDP) backbone to capture multi-level representations. It couples multi-channel inputs with deep stackable RNN layers and a pooling-based integration strategy, augmented by a directionality-based data augmentation method that generates inverse relations to mitigate data sparseness. On SemEval-2010 Task 8, DRNNs with augmentation achieve $F_1$ scores of $86.1\%$, surpassing prior state-of-the-art, with depth-4 networks offering the best performance. The approach demonstrates how deep temporal modeling and principled data augmentation can significantly improve relation classification in data-scarce settings, with potential impact on NLP pipelines relying on relation extraction.

Abstract

Nowadays, neural networks play an important role in the task of relation classification. By designing different neural architectures, researchers have improved the performance to a large extent in comparison with traditional methods. However, existing neural networks for relation classification are usually of shallow architectures (e.g., one-layer convolutional neural networks or recurrent networks). They may fail to explore the potential representation space in different abstraction levels. In this paper, we propose deep recurrent neural networks (DRNNs) for relation classification to tackle this challenge. Further, we propose a data augmentation method by leveraging the directionality of relations. We evaluated our DRNNs on the SemEval-2010 Task~8, and achieve an F1-score of 86.1%, outperforming previous state-of-the-art recorded results.

Improved Relation Classification by Deep Recurrent Neural Networks with Data Augmentation

TL;DR

This work addresses relation classification by introducing Deep Recurrent Neural Networks (DRNNs) built on the Shortest Dependency Path (SDP) backbone to capture multi-level representations. It couples multi-channel inputs with deep stackable RNN layers and a pooling-based integration strategy, augmented by a directionality-based data augmentation method that generates inverse relations to mitigate data sparseness. On SemEval-2010 Task 8, DRNNs with augmentation achieve scores of , surpassing prior state-of-the-art, with depth-4 networks offering the best performance. The approach demonstrates how deep temporal modeling and principled data augmentation can significantly improve relation classification in data-scarce settings, with potential impact on NLP pipelines relying on relation extraction.

Abstract

Nowadays, neural networks play an important role in the task of relation classification. By designing different neural architectures, researchers have improved the performance to a large extent in comparison with traditional methods. However, existing neural networks for relation classification are usually of shallow architectures (e.g., one-layer convolutional neural networks or recurrent networks). They may fail to explore the potential representation space in different abstraction levels. In this paper, we propose deep recurrent neural networks (DRNNs) for relation classification to tackle this challenge. Further, we propose a data augmentation method by leveraging the directionality of relations. We evaluated our DRNNs on the SemEval-2010 Task~8, and achieve an F1-score of 86.1%, outperforming previous state-of-the-art recorded results.

Paper Structure

This paper contains 16 sections, 3 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. The Proposed Methodology
  4. Overview
  5. Recurrent Neural Networks on Shortest Dependency Path
  6. Deep Recurrent Neural Networks
  7. Data Augmentation
  8. Training Objective
  9. Experiments
  10. Dataset
  11. Hyperparameter Settings
  12. Data Augmentation Details
  13. RNNs vs. CNNs
  14. Overall Performance
  15. Analysis of DRNNs' Depth
  16. ...and 1 more sections

Figures (3)

  • Figure 1: (a) The dependency parse tree corresponding to the sentence "Jewelry and other smaller [valuables]$_{e_1}$ were locked in a [safe]$_{e_2}$ or a closet with a deadbolt." Red arrows indicate the shortest dependency path between $e_1$ and $e_2$. (b) The augmented data sample.
  • Figure 2: The overall architecture of DRNNs. Two recurrent neural networks pick up information along the shortest dependency path, separated by its common ancestor. We use four information channels, namely words, part-of-speech tags, grammatical relations (GR), and WordNet hypernyms.
  • Figure 3: Visualization of information propagation along multiple RNN layers.