WiNER: A Wikipedia Annotated Corpus for Named Entity Recognition

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 413–422, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP WiNER: A Wikipedia Annotated Corpus for Named Entity Recognition Abbas Ghaddar RALI-DIRO Universit´e de Montr´eal Montr´eal, Canada abbas.ghaddar@umontreal.ca Philippe Langlais RALI-DIRO Universit´e de Montr´eal Montr´eal, Canada felipe@iro.umontreal.ca Abstract We revisit the idea of mining Wikipedia in order to generate named-entity anno- tations. We propose a new methodology that we applied to the English Wikipedia to build WiNER, a large, high quality, an- notated corpus. We evaluate its useful- ness on 6 NER tasks, comparing 4 popular state-of-the art approaches. We show that LSTM-CRF is the approach that benefits the most from our corpus. We report im- pressive gains with this model when using a small portion of WiNER on top of the CONLL training material. Last, we pro- pose a simple but efficient method for ex- ploiting the full range of WiNER, leading to further improvements. 1 Introduction Named-Entity Recognition (NER) is the task of identifying textual mentions and classifying them into a predefined set of types. It is an important pre-processing step in NLP and Information Ex- traction. Various approaches have been proposed to tackle the task, including conditional random fields (Finkel et al., 2005), perceptrons (Ratinov and Roth, 2009), and neural network approaches (Collobert et al., 2011; Lample et al., 2016; Chiu and Nichols, 2016). One issue with NER is the small amount of annotated data available for training, and their limited scope (see Section 4.1). Furthermore, some studies (Onal and Karagoz, 2015; Augen- stein et al., 2017) have demonstrated that named- entity systems trained on news-wire data perform poorly when tested on other text genres. This mo- tivated some researchers to create a named-entity labelled corpus from Wikipedia. This was notably attempted by Nothman et al. (2008) and more re- cently revisited by Al-Rfou et al. (2015) in a mul- tilingual context. Both studies leverage the link structure of Wikipedia to generate named-entity annotations. Because only a tiny portion of texts in Wikipedia are anchored, some strategies are typ- ically needed to infer more annotations (Ghad- dar and Langlais, 2016b). Such a process typi- cally yields a noisy corpus for which filtering is required. In this paper, we revisit the idea of auto- matically extracting named-entity annotations out of Wikipedia. Similarly to the aforementioned works, we gather anchored strings in a page as well as their type according to Freebase (Bol- lacker et al., 2008) but, more importantly, we also generate annotations for texts not anchored in Wikipedia. We do this by considering coref- erence mentions of anchored strings as candidate annotations, and by exploiting the out-link struc- ture of Wikipedia. We applied our methodology on a 2013 English Wikipedia dump, leading to a large annotated corpus called WiNER, which con- tains more annotations than similar corpora and, as we shall see, is more useful for training NER systems. We discuss related work in Section 2 and present the methodology we used to …

Content truncated to 500 words