A data-centric way to improve entity linking in knowledge-based question answering.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Information Retrieval

Background:

• Entity linking is crucial for knowledge-based question answering (KBQA), mapping text mentions to knowledge base entities.
• Existing entity linking research primarily focuses on long texts, leaving a gap in short-text scenarios common in open-domain KBQA.
• Current neural network models often require structural adjustments and show limited performance across diverse datasets.

Purpose of the Study:

• To improve entity linking performance specifically for short texts in open-domain KBQA.
• To develop a data-centric approach that enhances existing entity linking models without structural modifications.
• To introduce a novel method for extracting and integrating domain-specific information from short texts.

Main Methods:

• Developed Domain information Mining and Explicit expressing (DME) to extract and append domain information to training data.
• Implemented a novel negative sampling approach to increase model robustness.
• Trained and evaluated entity linking models using DME-processed data on the KgCLUE benchmark.

Main Results:

• DME-processed data significantly improved the performance of baseline entity linking models.
• The approach enhanced entity linking without requiring changes to the underlying model architecture.
• Experimental results demonstrated the transferability of the DME approach to other datasets.

Conclusions:

• The DME approach offers an effective data-centric strategy for improving short-text entity linking in KBQA.
• This method enhances model performance and robustness, particularly in domain-specific contexts.
• The proposed technique is adaptable and shows promise for broader application in natural language understanding tasks.