Metaknowledge Enhanced Open Domain Question Answering with Wiki Documents

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Knowledge Representation

Background:

• Traditional triplet-based knowledge bases struggle with loose associations and weak logic in open-domain question answering.
• Existing methods face challenges in effectively structuring and utilizing knowledge for complex queries.

Purpose of the Study:

• To propose a novel metaknowledge-enhanced approach for open-domain question answering.
• To overcome the limitations of current knowledge representation methods in large-scale knowledge bases.

Main Methods:

• Developed an automatic approach to extract metaknowledge and construct a metaknowledge network from Wiki documents.
• Introduced GE4MK, an original graph encoder, to model the metaknowledge network with hierarchical and semantic features.
• Proposed MEGr-Net, a metaknowledge-enhanced graph reasoning model, integrating relational and neighboring interactions.

Main Results:

• Experimental results demonstrate the superiority of metaknowledge over mainstream triplet-based knowledge.
• MEGr-Net shows significant improvements in open-domain question answering tasks.
• The study confirms the influence of graph reasoning and pre-trained language models on metaknowledge-enhanced approaches.

Conclusions:

• Metaknowledge provides a more robust and structured representation for question answering.
• The proposed GE4MK encoder and MEGr-Net model offer an effective solution for enhancing open-domain question answering.
• Future research can further explore the integration of graph reasoning and pre-trained language models with metaknowledge.