Abstract: Belinda Z. Li, Sewon Min, Srinivasan Iyer, Yashar Mehdad, Wen-tau Yih. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).
2020.
Introduction
Entity Linking (EL) is the critical process of identifying text spans (mentions) and mapping them to unique entries in a knowledge base, such as Wikipedia. While EL is well-studied for long-form text like news articles, it remains a significant challenge for short, user-generated questions. Questions are often characterized by a lack of context, missing capitalization, and ambiguous phrasing (e.g., "shaq" referring to Shaquille O'Neal).
Traditional EL systems often rely on complex multi-stage pipelines that separate mention detection from entity disambiguation. These pipelines are not only slow but also suffer from error propagation. High-performance models like BLINK frequently require pre-defined "gold" mention boundaries, making them impractical for real-world applications where boundaries are unknown. To address these issues, the authors propose ELQ (Efficient Entity Linking for Questions), an end-to-end model designed for speed and accuracy in one pass.
Methodology
ELQ utilizes a BERT-based bi-encoder architecture to project both questions and entities into a shared high-dimensional vector space. The system is designed for maximum efficiency by pre-computing entity representations and performing joint detection and linking.
The Bi-Encoder Architecture
- Entity Encoder: This component processes Wikipedia titles and short descriptions to create fixed embeddings for every entity. These are stored in a fast, searchable index (e.g., FAISS), allowing for rapid retrieval during inference.
- Question Encoder: A BERT-based encoder produces token-level representations for the input question.
- Joint Mention Detection and Linking: Unlike pipeline systems, ELQ performs both tasks simultaneously. It uses token embeddings to predict the start and end of potential entity spans. For each candidate span, the model computes an average embedding and performs a Maximum Inner Product Search (MIPS) against the pre-computed entity index.
By scoring all possible spans in a single forward pass of BERT, ELQ avoids the computational bottleneck of multiple passes or separate candidate generation steps.
Experimental Results
The researchers evaluated ELQ on several benchmarks, including WebQSP_EL and GraphQ_EL. The results demonstrate that ELQ significantly outperforms existing state-of-the-art (SOTA) models in both accuracy and speed.
Accuracy and Generalization
On the WebQSP_EL benchmark, ELQ achieved an F1 score of 89.0, a substantial improvement over the previous SOTA (VCG) which scored 74.7. Furthermore, ELQ showed strong zero-shot capabilities; when trained on Wikipedia and tested on GraphQ_EL without any domain-specific training, it still outperformed competing models.
Inference Speed
Efficiency is a core strength of ELQ. It processes approximately 1.57 questions per second on a single CPU, making it roughly twice as fast as the BLINK-based pipeline (0.80 Q/s) and over three times faster than VCG (0.45 Q/s). This speedup is primarily due to the single-pass architecture, which minimizes the overhead of the BERT forward pass.
Discussion and Impact
The study highlights that joint training of mention detection and entity disambiguation is mutually beneficial; the model learns to identify mentions more effectively by understanding which entities actually exist in the knowledge base. When integrated into downstream Question Answering systems like GraphRetriever, ELQ improved end-to-end answer accuracy by up to 6%.
While ELQ is highly effective, qualitative analysis revealed that many remaining errors are due to inherent ambiguity in short questions (e.g., distinguishing between "Washington state" and "Washington D.C.") or insufficient context. Future work may involve incorporating broader session context or adapting the architecture for longer documents where the computational cost of scoring every span must be managed.
Summary updated