DFKI: Multi-objective Optimization for the Joint Disambiguation of Entities and Nouns & Deep Verb Sense Disambiguation

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 335–339, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics DFKI: Multi-objective Optimization for the Joint Disambiguation of Entities and Nouns & Deep Verb Sense Disambiguation Dirk Weissenborn LT, DFKI Alt-Moabit 91c Berlin, Germany dirk.weissenborn@dfki.de Feiyu Xu LT, DFKI Alt-Moabit 91c Berlin, Germany feiyu@dfki.de Hans Uszkoreit LT, DFKI Alt-Moabit 91c Berlin, Germany uszkoreit@dfki.de Abstract We introduce an approach to word sense dis- ambiguation and entity linking that combines a set of complementary objectives in an exten- sible multi-objective formalism. During dis- ambiguation the system performs continuous optimization to find optimal probability dis- tributions over candidate senses. Verb senses are disambiguated using a separate neural net- work model. Our results on noun and verb sense disambiguation as well as entity linking outperform all other submissions on the Se- mEval 2015 Task 13 for English. 1 Introduction The task of assigning the correct meaning to a given word or entity mention in a document is called word sense disambiguation (WSD) (Navigli, 2009) or en- tity linking (EL) (Bunescu and Pasca, 2006), respec- tively. Successful disambiguation requires not only an understanding of the topic or domain a document is dealing with (global), but also an analysis of how an individual word is used within its local context. E.g., the meanings of the word “newspaper” as the company or the physical product, often cannot be distinguished by the topic, but by recognizing which type of meaning fits best into the local context of its occurrence. On the other hand, for an ambiguous entity mention such as “Michael Jordan” it is impor- tant to recognize the topic of the wider context to distinguish, e.g., between the basketball player and the machine learning expert. The combination of the two most commonly used reference knowledge bases for WSD and EL, e.g., WordNet (Fellbaum, 1998) and Wikipedia, by Ba- belNet (Navigli and Ponzetto, 2012) has enabled a new line of research towards the joint disambigua- tion of words and named entities. Babelfy (Moro et al., 2014) has shown the potential of combining these two tasks in a purely knowledge-driven ap- proach that jointly finds connections between po- tential word senses in the global context. On the other hand, typical supervised methods (Zhong and Ng, 2010) trained on sense-annotated corpora are usually quite successful in dealing with individual words in a local context. Hoffart et al. (2011) rec- ognize the importance of combining both local con- text and global context for robust disambiguation. However, their approach is limited to EL, where op- timization is performed in a discrete setting. We present a system that combines disambigua- tion objectives for both global and local contexts into a single multi-objective function. In contrast to prior work we model the problem in a continuous setting based on probability distributions over can- didate meanings. Our approach exploits lexical and encyclopedic knowledge, local context information and statistics of the mapping from text to candidate meanings. Furthermore, we introduce a deep learn- …

Content truncated to 500 words