Abstract
Information extraction from microblog posts is an important task, as today microblogs capture an unprecedented amount of information and provide a view into the pulse of the world. As the core component of information extraction, we consider the task of Twitter entity linking in this paper. In the current entity linking literature, mention detection and entity disambiguation are frequently cast as equally important but distinct problems. However, in our task, we find that mention detection is often the performance bottleneck. The reason is that messages on micro-blogs are short, noisy and informal texts with little context, and often contain phrases with ambiguous meanings. To rigorously address the Twitter entity linking problem, we propose a structural SVM algorithm for entity linking that jointly optimizes mention detection and entity disambiguation as a single end-to-end task. By combining structural learning and a variety of first-order, second-order, and context-sensitive features, our system is able to outperform existing state-of-the art entity linking systems by 15% F1.
Introduction
Entity Linking (EL) involves identifying a phrase in text (a "mention") and connecting it to a specific entry in a knowledge base like Wikipedia. While this is relatively straightforward in well-formatted news articles, Twitter presents a unique challenge. Tweets are short, riddled with typos, lack proper capitalization, and use informal slang.
Traditionally, EL systems use a two-step pipeline: mention detection followed by entity disambiguation. However, this study identifies that in the noisy world of microblogs, mention detection is the primary bottleneck. As shown in the data statistics, if a system correctly identifies the boundaries of a mention, simply selecting the most popular matching Wikipedia page (P@1) yields an accuracy of over 85% across datasets. The real challenge lies in distinguishing whether a phrase like "one day" is a movie title or common parlance.
Methodology
The proposed solution is a Structural SVM (SSVM) framework designed for "end-to-end" linking. Instead of a sequential pipeline, the model considers all possible segmentation and linking assignments simultaneously to find the most globally consistent interpretation.
Joint Optimization and Feature Engineering
The model utilizes "entity-specific" features, such as Wikipedia popularity, to inform the initial detection of mentions. The feature set includes:
- First-order features: Link probability, capitalization rates, and entity types.
- Second-order features: Jaccard similarity between Wikipedia pages to ensure semantic cohesiveness (e.g., ensuring "Ben Affleck" and "Gigli" appear in a related context).
- Context Mining: The authors mined "contextual words" from 20 million unlabeled tweets to identify informal signals. For instance, slang like "nw" (now watching) serves as a strong indicator that the following text is a media entity.
Because joint optimization with second-order features is NP-hard, the system employs a beam search algorithm for efficient inference.
Experimental Results
The SSVM system was evaluated against three datasets: a training set and two test sets (one based on the Ritter dataset and one focused on entertainment entities). The results demonstrate that the SSVM significantly outperforms existing state-of-the-art tools like Cucerzan and TagMe.
On Test 1, the SSVM achieved an $F_1$ score of 68.0%, a substantial lead over Cucerzan (51.1%) and TagMe (49.7%). This performance was maintained on the entertainment-focused Test 2 (65.2% $F_1$), proving the model's robustness across different domains. The inclusion of second-order "cohesiveness" features and mined contextual slang provided consistent incremental gains.
Conclusion
This study highlights that for microblogging platforms, the difficulty of entity linking lies in detection rather than disambiguation. By treating these as a joint task through a Structural SVM, the researchers achieved a 15% improvement in $F_1$ score over previous methods. Future work may address the computational complexity of second-order features and expand the system to "Open World" entity linking for entities not yet present in Wikipedia.
Summary updated