Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
102 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Reinforcement Learning for Relation Classification from Noisy Data (1808.08013v1)

Published 24 Aug 2018 in cs.IR, cs.LG, and stat.ML

Abstract: Existing relation classification methods that rely on distant supervision assume that a bag of sentences mentioning an entity pair are all describing a relation for the entity pair. Such methods, performing classification at the bag level, cannot identify the mapping between a relation and a sentence, and largely suffers from the noisy labeling problem. In this paper, we propose a novel model for relation classification at the sentence level from noisy data. The model has two modules: an instance selector and a relation classifier. The instance selector chooses high-quality sentences with reinforcement learning and feeds the selected sentences into the relation classifier, and the relation classifier makes sentence level prediction and provides rewards to the instance selector. The two modules are trained jointly to optimize the instance selection and relation classification processes. Experiment results show that our model can deal with the noise of data effectively and obtains better performance for relation classification at the sentence level.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Jun Feng (55 papers)
  2. Minlie Huang (226 papers)
  3. Li Zhao (150 papers)
  4. Yang Yang (884 papers)
  5. Xiaoyan Zhu (54 papers)
Citations (335)

Summary

Reinforcement Learning for Relation Classification from Noisy Data

This paper addresses the challenge of relation classification in NLP with a particular focus on handling noisy data arising from distant supervision. Unlike traditional relation classification which often suffers from noisy labeling, the authors introduce a novel framework utilizing reinforcement learning to address this prevalent issue. The proposed model innovatively separates the relation classification task into two modules: an instance selector and a relation classifier, each trained to work synergistically.

The instance selector employs reinforcement learning to filter noisy data by identifying and selecting high-quality sentence instances that truly describe the relations. This process leverages a policy network to evaluate each sentence and decide whether it should be included in the cleansed training set. The decision is based on delayed rewards calculated by the relation classifier, which assesses the utility of the selected sentences. This mechanism allows the instance selector to learn from indirect supervision and iteratively improve its selection policy, unlike direct approaches that require precise sentence-level annotations.

The relation classifier, on the other hand, is built on a convolutional neural network (CNN) architecture. It performs sentence-level relation classification using the data curated by the instance selector. The CNN model is trained to maximize the likelihood of relations based on sentence content, conducting the classification with significantly reduced noise. This results in more reliable relation predictions.

Experiments demonstrate the efficacy of this framework. The authors show robust numerical improvements over baseline models, including traditional CNNs and attention-based approaches such as CNN+Max and CNN+ATT. The instance selector's strategy allows a more precise mapping of relations to specific sentences, leading to enhanced sentence-level prediction accuracy. Notably, their model shows improvement in macro F1-score and accuracy compared to previous methods, underscoring its capability to better handle noisy data.

The implications of this research extend beyond mere performance enhancement in relation classification. By decoupling instance selection from classification and employing reinforcement learning, the approach advances the methodology for handling noisy data in NLP. This could precipitate improved models for other NLP tasks that rely on noisy datasets, such as sentiment classification or entity extraction.

Looking toward the future, refining the reward mechanism and exploring alternative reinforcement learning strategies could further bolster the model's effectiveness. Additionally, expanding the framework to integrate different neural architectures or feature representations may yield further enhancements. This approach, by prioritizing high-quality instance selection, sets a precedent for effectively managing noisy data, potentially influencing broader applications in artificial intelligence that deal with imperfect data sources.