Novel knowledge-based system with relation detection and textual evidence for question answering research

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Information Retrieval

Background:

• Knowledge-based question answering (KBQA) is a significant research area, crucial for extracting information from large-scale knowledge bases (KBs).
• Existing KBQA systems often use limited shallow methods for relation detection, failing to capture deep semantic meaning.
• Knowledge bases are inherently incomplete, necessitating strategies to incorporate external unstructured information.

Purpose of the Study:

• To develop a novel KBQA system, QARDTE, that improves relation detection and compensates for KB incompleteness.
• To address the semantic gap in relation detection using deep learning models.
• To integrate unstructured textual evidence into the KBQA process for enhanced accuracy.

Main Methods:

• Utilized bidirectional long-short term memory networks (Bi-LSTMs) at multiple abstraction levels for robust relation detection.
• Incorporated external unstructured text to extract supporting evidence for question answering.
• Developed a re-ranking process combining KB relation information with extracted textual evidence.

Main Results:

• Achieved state-of-the-art performance on two benchmark datasets, demonstrating significant improvements over existing KBQA systems.
• The QARDTE system attained F1 scores of 0.558 (+2.8%) and 0.663 (+5.7%).
• The model showed robustness against diverse linguistic expressions and complex questions involving multiple relations.

Conclusions:

• The proposed QARDTE method effectively enhances relation detection through deep semantic representations.
• Leveraging external unstructured text significantly compensates for the limitations of incomplete knowledge bases.
• The QARDTE system represents a substantial advancement in the field of knowledge-based question answering.