Language-based reasoning graph neural network for commonsense question answering.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Language models (LMs) are crucial for common-sense reasoning in question answering (CSQA).
• Increasing LM parameters yields diminishing returns; external knowledge integration is key.
• Graph neural networks (GNNs) enhance performance but struggle with diverse knowledge sources and text-knowledge context.

Purpose of the Study:

• To propose Language-Based Reasoning Graph Neural Network (LBR-GNN) for improved CSQA.
• To address challenges in exploiting diverse knowledge and capturing text-knowledge context.
• To enhance common-sense understanding and reasoning capabilities.

Main Methods:

• Representing questions, answers, and external knowledge using a language model.
• Regulating external knowledge into a consistent textual form and encoding it with an LM.
• Building a GNN with language-level edge representations and a novel edge aggregation method for GNN updates and reasoning.

Main Results:

• LBR-GNN demonstrated a performance boost of over 5% on the CommonsenseQA dataset compared to state-of-the-art methods.
• The method achieved this improvement with a comparable number of additional parameters.
• Effective performance was also observed on CommonsenseQA-IH and OpenBookQA datasets.

Conclusions:

• LBR-GNN effectively integrates external knowledge for enhanced common-sense question answering.
• The proposed language-based GNN approach successfully captures contextual information between text and knowledge.
• LBR-GNN offers a promising direction for advancing AI reasoning capabilities.