Utilizing Entity-Based Gated Convolution and Multilevel Sentence Attention to Improve Distantly Supervised Relation...

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Distant supervision is crucial for large-scale dataset creation in relation extraction (RE).
• Automatically generated datasets often contain intrasentence noise and mislabeled sentences, hindering RE performance.
• Existing methods struggle to effectively mitigate these noise types.

Purpose of the Study:

• To propose a novel distantly supervised relation extraction model that addresses intrasentence noise and wrongly labeled sentences.
• To enhance the feature extraction and sentence selection capabilities in distantly supervised RE.
• To achieve state-of-the-art performance in relation extraction tasks.

Main Methods:

• An entity-based gated convolution sentence encoder is employed to extract entity-pair-specific features and filter noise.
• A multilevel sentence selective attention (Matt) module fuses bag information for fine-grained query vectors.
• The Matt module aids in identifying valid sentences and reducing the impact of noisy data.

Main Results:

• The proposed model effectively reduces the influence of both intrasentence noise and wrongly labeled sentences.
• Experimental results on a large-scale benchmark dataset demonstrate superior performance.
• The model achieves state-of-the-art results in distantly supervised relation extraction.

Conclusions:

• The novel model successfully tackles noise challenges in distantly supervised relation extraction.
• The entity-based gated convolution and multilevel attention mechanisms are key to the model's effectiveness.
• This work advances the field of relation extraction by providing a robust solution for noisy datasets.