Entity Disambiguation by Knowledge and Text Jointly Embedding

Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pages 260–269, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Entity Disambiguation by Knowledge and Text Jointly Embedding Wei Fang1,2, Jianwen Zhang3, Dilin Wang4, Zheng Chen3, Ming Li1,2 1SYSU-CMU Joint Institute of Engineering, School of Electronics and Information Technology, Sun Yat-Sen University 2SYSU-CMU Shunde International Joint Research Institute {fangwei7@mail2, liming46@mail}.sysu.edu.cn 3Microsoft, Redmond, WA {jiazhan, zhengc}@microsoft.com 4Computer Science, Dartmouth College dilin.wang.gr@dartmouth.edu Abstract For most entity disambiguation systems, the secret recipes are feature representa- tions for mentions and entities, most of which are based on Bag-of-Words (BoW) representations. several drawbacks:Commonly, BoW has (1) It ignores the intrinsic meaning of words/entities; (2) It often results in high-dimension vector spaces and expensive computation; (3) For different applications, methods of design- ing handcrafted representations may be quite different, lacking of a general guide- line. In this paper, we propose a different approach named EDKate. We first learn low-dimensional continuous vector repre- sentations for entities and words by jointly embedding knowledge base and text in the same vector space. Then we utilize these embeddings to design simple but effec- tive features and build a two-layer disam- biguation model. Extensive experiments on real-world data sets show that (1) The embedding-based features are very effec- tive. Even a single one embedding-based feature can beat the combination of sev- eral BoW-based features. (2) The superi- ority is even more promising in a difficult set where the mention-entity prior cannot work well. (3) The proposed embedding method is much better than trivial imple- mentations of some off-the-shelf embed- ding algorithms. (4) We compared our EDKate with existing methods/systems and the results are also positive. 1 Introduction Entity disambiguation is the task of linking entity mentions in unstructured text to the correspond- ing entities in a knowledge base. For example, in the sentence “Michael Jordan is newly elected as AAAI fellow”, the mention “Michael Jordan” should be linked to “Michael I. Jordan” (Berke- ley Professor) rather than “Michael Jordan” (NBA Player). Formally, given a set of mentions M = {m1, m2, ..., mk} (specified or detected automati- cally) in a document d, for each mention mi ∈ M, the task is to find the correct entity ei in the knowl- edge base (KB) K to which the mention mi refers. There are various methods proposed for the problem in the past decades. But generally speaking, an entity disambiguation method is commonly composed of three stages/components. (1) Constructing representations for men- tions/entities from raw data, often as the form of sparse vectors. (2) Extracting features for disambiguation models based on the represen- tations of mentions and entities constructed in stage (1). (3) Optimizing the disambiguation model by empirically setting or learning weights on the extracted features, e.g., by training a classifier/ranker. There exist few features directly defined by heuristics, skipping the first stage. For example, string similarity or edit distance between a mention surface and an entity’s canonical form (Cucerzan, 2011; Cassidy et al., 2011), and the prior …

Content truncated to 500 words