Abstract
Entity Linking (EL) is a fundamental NLP task that maps textual mentions to specific entries in a knowledge base. While well-studied in formal texts, EL remains challenging in Community Question Answering (CQA) environments due to short contexts and informal language. This paper proposes a novel framework that leverages auxiliary data—such as topic tags, user history, and parallel answers—to disambiguate mentions. Our approach achieves a state-of-the-art accuracy of 92.02% on the newly introduced QuoraEL dataset.
Introduction
Standard Entity Linking models often struggle with the informal and concise nature of CQA platforms like Quora or Stack Overflow. Mentions like "Roosevelt" or "AGI" are difficult to resolve when the immediate text provides minimal context. However, CQA platforms provide unique metadata that can serve as "side information." For example, a post tagged with "World War II" strongly suggests "Franklin D. Roosevelt" over "Theodore Roosevelt."
The authors formalize this as the CQAEL task, focusing on harnessing these auxiliary signals to improve linking accuracy in noisy, low-context environments.
Proposed Methodology: A Dual-Module Framework
The researchers developed a framework consisting of two primary components designed to handle both local context and external metadata.
1. Base Module
The Base Module utilizes an XLNet-based cross-encoder to model the relationship between a mention's local context and the candidate entity's description. It also incorporates a "prior probability" score, representing the global popularity of an entity for a given mention string.
2. Auxiliary Data Module
The core innovation lies in the Auxiliary Data Module, which processes potentially voluminous and noisy metadata through a two-stage pipeline:
- Selection: Using string similarity metrics (e.g., Levenshtein, Jaro-Winkler), the system identifies the top-$k$ most relevant auxiliary texts from the available pool. The study found that $k=3$ provides the optimal balance between signal and noise.
- Encoding: A Longformer model jointly encodes the entity description alongside the selected auxiliary texts. Longformer's ability to process long sequences makes it ideal for handling multiple context snippets.
The final entity ranking score is a fusion of the outputs from both modules, trained using cross-entropy loss.
Experimental Evaluation
To evaluate the framework, the authors introduced QuoraEL, a dataset containing over 8,000 labeled mentions across 504 CQA texts. The knowledge base was derived from a 2019 Wikipedia dump.
Key Results
- State-of-the-Art Performance: The full framework reached 92.02% accuracy, outperforming established models like BLINK (87.92%) and Zeshel (88.91%) even when those models were provided with auxiliary data.
- Ablation Insights: Removing auxiliary data led to a 2.65% drop in accuracy, confirming the value of external metadata. Parallel answers and topic tags were found to be particularly potent signals.
- Model Agnosticism: The Auxiliary Data Module can be integrated into existing EL models. When added to the Deep-ED model, it yielded a significant 5.6% performance boost.
Conclusion
By effectively filtering and encoding auxiliary data, this research demonstrates that the "noise" of CQA platforms can be transformed into a powerful asset for entity disambiguation. While computationally more expensive than simple bi-encoders, the performance gains justify the use of cross-encoders and Longformers in this domain. Future work may involve extending this framework to more technical domains, such as software engineering forums.
Summary updated