Named Entity Disambiguation for Noisy Text

Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 58–68, Vancouver, Canada, August 3 - August 4, 2017. c⃝2017 Association for Computational Linguistics Named Entity Disambiguation for Noisy Text Yotam Eshel1 Noam Cohen1 Kira Radinsky1,2 Shaul Markovitch1 Ikuda Yamada3 Omer Levy4 1Technion - Israel Institute of Technology, Haifa, Israel 2eBay Research, Israel 3Studio Ousia, Fujisawa, Kanagawa, Japan 4University of Washington, Seattle, WA Abstract We address the task of Named Entity Disambiguation (NED) for noisy text. We present WikilinksNED, a large-scale NED dataset of text fragments from the web, which is significantly noisier and more challenging than existing news- based datasets. To capture the limited and noisy local context surrounding each mention, we design a neural model and train it with a novel method for sam- pling informative negative examples. We also describe a new way of initializing word and entity embeddings that signifi- cantly improves performance. Our model significantly outperforms existing state-of- the-art methods on WikilinksNED while achieving comparable performance on a smaller newswire dataset. 1 Introduction Named Entity Disambiguation (NED) is the task of linking mentions of entities in text to a given knowledge base, such as Freebase or Wikipedia. NED is a key component in Entity Linking (EL) systems, focusing on the disambiguation task it- self, independently from the tasks of Named En- tity Recognition (detecting mention bounds) and Candidate Generation (retrieving the set of poten- tial candidate entities). NED has been recognized as an important component in NLP tasks such as semantic parsing (Berant and Liang, 2014). Current research on NED is mostly driven by a number of standard datasets, such as CoNLL- YAGO (Hoffart et al., 2011), TAC KBP (Ji et al., 2010) and ACE (Bentivogli et al., 2010). These datasets are based on news corpora and Wikipedia, which are naturally coherent, well-structured, and rich in context. Global disambiguation models (Guo and Barbosa, 2014; Pershina et al., 2015; Globerson et al., 2016) leverage this coherency by jointly disambiguating all the mentions in a sin- gle document. However, domains such as web- page fragments, social media, or search queries, are often short, noisy, and less coherent; such do- mains lack the necessary contextual information for global methods to pay off, and present a more challenging setting in general. In this work, we investigate the task of NED in a setting where only local and noisy context is available. In particular, we create a dataset of 3.2M short text fragments extracted from web pages, each containing a mention of a named en- tity. Our dataset is far larger than previously col- lected datasets, and contains 18K unique mentions linking to over 100K unique entities. We have em- pirically found it to be noisier and more challeng- ing than existing datasets. For example: “I had no choice but to experiment with other indoor games. I was born in At- lantic City so the obvious next choice was Monopoly. I played until I became a successful Captain of Industry.” This short fragment is considerably less structured …

Content truncated to 500 words