Discourse-Aware Graph Networks for Textual Logical Reasoning.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Textual logical reasoning is crucial for question-answering (QA).
• Current QA systems often overlook passage-level logical structures, focusing on entity relations.
• Logical structures like entailment and contradiction are underexplored in QA.

Purpose of the Study:

• To propose logic structural-constraint modeling for logical reasoning QA.
• To introduce discourse-aware graph networks (DAGNs) for enhanced logical reasoning.
• To improve answer prediction by incorporating logical structure features.

Main Methods:

• Constructing logic graphs using discourse connectives and logic theories.
• Employing an edge-reasoning mechanism for end-to-end learning of logic relations.
• Integrating learned logic features with fundamental features in a general encoder for answer prediction.

Main Results:

• DAGNs effectively build and represent logical structures in text.
• Learned logic features significantly improve performance on textual logical reasoning QA tasks.
• Demonstrated strong zero-shot transfer capabilities to unseen logical texts.

Conclusions:

• The proposed DAGNs are effective for modeling logical structures in textual reasoning.
• The learned logic features generalize well, enhancing QA performance across datasets.
• Highlights the importance of explicit logical structure modeling in NLP.