Transforming Wikipedia into a Large-Scale Fine-Grained Entity Type Corpus

Transforming Wikipedia into a Large-Scale Fine-Grained Entity Type Corpus Abbas Ghaddar, Philippe Langlais RALI-DIRO Montreal, Canada abbas.ghaddar@umontreal.ca, felipe@iro.umontreal.ca Abstract This paper presents WiFiNE, an English corpus annotated with fine-grained entity types. We propose simple but effective heuristics we applied to English Wikipedia to build a large, high quality, annotated corpus. We evaluate the impact of our corpus on the fine-grained entity typing system of Shimaoka et al. (2017), with 2 manually annotated benchmarks, FIGER (GOLD) and ONTONOTES. We report state-of-the-art performances, with a gain of 0.8 micro F1 score on the former dataset and a gain of 2.7 macro F1 score on the latter one, despite the fact that we employ the same quantity of training data used in previous works. We make our corpus available as a resource for future works. Keywords: Annotated Corpus, Fine-Grained Entity Type, Wikipedia 1. Introduction Entity typing is the task of classifying textual mentions into their respective types. While the standard Named-Entity Recognition (NER) task focuses on a small set of types (e.g. 4 classes defined by the CONLL shared task-2003 (Tjong Kim Sang and De Meulder, 2003)), fine-grained tagging deals with much larger type sets (e.g. 112 types used in (Ling and Weld, 2012)). Entity typing has received an increasing interest lately from the NLP community, due its role in Relation Extraction, Entity Linking, Question An- swering, etc. One issue in fine-grained typing is the absence of a well- established training corpus. The large number of types makes it difficult to manually annotate the amount of data needed for training. This bottleneck was addressed by us- ing an automatic annotation procedure (Section 2), which follows two steps: 1. Identifying and linking entity mentions to a Knowl- edge Base (typically Freebase). 2. Assigning to each mention the set of types that apply in the context of the sentence. Step 1 suffers a number of issues: lack of coverage when Wikipedia is used as a source (Ghaddar and Langlais, 2016b), and entity linking which is error prone (Ren et al., 2016). Step 2 also has limitations: the type of a mention is often resolved with strict pruning heuristics (regardless of the context) as in (Gillick et al., 2014); or the type of a mention is kept ambiguous following (Shimaoka et al., 2017). For instance, in the sentence: “Gonzales embarked on a pop career as the leader of the alternative rock band Son.” The entity Chilly Gonzales has 3 labels in Freebase: musician, writer , actor but only musician applies here. In this paper, we revisit the idea of automatically extract- ing fine-grained entity annotations from Wikipedia. Sim- ilarly to previous works, we gather annotations from an- chored texts in an article, as well as their associated types in Freebase (Bollacker et al., 2008). In addition, we also generate annotations for texts not anchored in Wikipedia following (Ghaddar and Langlais, 2017). We do this by considering coreference mentions of anchored texts as can- didate annotations, and by exploiting the out-link structure of Wikipedia. We propose an …

Content truncated to 500 words