Efficient Collective Entity Linking with Stacking

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 426–435, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Efficient Collective Entity Linking with Stacking Zhengyan He† Shujie Liu‡ Yang Song† Mu Li‡ Ming Zhou‡ Houfeng Wang†∗ † Key Laboratory of Computational Linguistics (Peking University) Ministry of Education,China ‡ Microsoft Research Asia hezhengyan.hit@gmail.com {shujliu,muli,mingzhou}@microsoft.com songyangmagic@gmail.com wanghf@pku.edu.cn Abstract Entity disambiguation works by linking am- biguous mentions in text to their correspond- ing real-world entities in knowledge base. Re- cent collective disambiguation methods en- force coherence among contextual decisions at the cost of non-trivial inference processes. We propose a fast collective disambiguation approach based on stacking. First, we train a local predictor g0 with learning to rank as base learner, to generate initial ranking list of can- didates. Second, top k candidates of related instances are searched for constructing expres- sive global coherence features. A global pre- dictor g1 is trained in the augmented feature space and stacking is employed to tackle the train/test mismatch problem. The proposed method is fast and easy to implement. Exper- iments show its effectiveness over various al- gorithms on several public datasets. By learn- ing a rich semantic relatedness measure be- tween entity categories and context document, performance is further improved. 1 Introduction When extracting knowledge from natural language text into a machine readable format, ambiguous names must be resolved in order to tell which real- world entity the name refers to. The task of linking names to knowledge base is known as entity linking or disambiguation (Ji et al., 2011). The resulting text is populated with semantic rich links to knowledge base like Wikipedia, and ready for various down- stream NLP applications. ∗Corresponding author Previous researches have proposed several kinds of effective approaches for this problem. Learning to rank (L2R) approaches use hand-crafted features f(d, e) to describe the similarity or dissimilarity be- e.tween contextual document d and entity definition L2R approaches are very flexible and expres- sive. Features like name matching, context similar- ity (Li et al., 2009; Zheng et al., 2010; Lehmann et al., 2010) and category context correlation (Bunescu and Pasca, 2006) can be incorporated with ease. Nevertheless, decisions are made independently and inconsistent results are found from time to time. Collective approaches utilize dependencies be- tween different decisions and resolve all ambiguous mentions within the same context simultaneously (Han et al., 2011; Hoffart et al., 2011; Kulkarni et al., 2009; Ratinov et al., 2011). Collective ap- proaches can improve performance when local ev- idence is not confident enough. They often utilize semantic relations across different mentions, and is why they are called global approaches, while L2R methods fall into local approaches (Ratinov et al., 2011). However, collective inference processes are often expensive and involve an exponential search space. We propose a collective entity linking method based on stacking. Stacked generalization (Wolpert, 1992) is a powerful meta learning algorithm that uses two levels of learners. The predictions of the first learner are taken as augmented features …

Content truncated to 500 words