BASHI: A Corpus of Wall Street Journal Articles Annotated with Bridging Links

BASHI: A Corpus of Wall Street Journal Articles Annotated with Bridging Links Ina R¨osiger Institute for Natural Language Processing University of Stuttgart, Germany Abstract This paper presents a corpus resource for the anaphoric phenomenon of bridging, named BASHI. The corpus consisting of 50 Wall Street Journal (WSJ) articles adds bridging anaphors and their antecedents to the other gold annotations that have been created as part of the OntoNotes project (Weischedel et al., 2011). Bridging anaphors are context-dependent expressions that do not refer to the same entity as their antecedent, but to a related entity. Bridging resolution is an under-researched area of NLP, where the lack of annotated training data makes the application of statistical models difficult. Thus, we believe that the corpus is a valuable resource for researchers interested in anaphoric phenomena going beyond coreference, as it can be combined with other corpora to create a larger corpus resource. The corpus contains 57,709 tokens and 459 bridging pairs and is available for download in an offset-based format and a CoNLL-12 style bridging column that can be merged with the other annotation layers in OntoNotes. The paper also reviews previous annotation efforts and different definitions of bridging and reports challenges with respect to the bridging annotation. Keywords: Corpus Resource, Bridging, Anaphora, Wall Street Journal, OntoNotes, English 1. Introduction Bridging is an anaphoric phenomenon where the interpre- tation of a bridging anaphor, sometimes also called associa- tive anaphor (Hawkins, 1978), is based on the non-identical associated antecedent. The associated NLP task of bridging resolution is about linking these anaphoric noun phrases and their antecedents, where both do not refer to the same referent, but are related in a way that is not explicitly stated. Bridging anaphors are thus discourse-new, but dependent on previous context. (1) I went to a wedding last weekend. The bride was a friend of mine.1. (2) What is the book about? The answer isn’t trivial. One can think about bridging anaphors as expressions with an implicit argument, e.g. the bride (at a wedding) or the answer (to this question). 1.1. Motivation Compared to coreference resolution, which has become one of the standard NLP tasks, with its own track at most NLP conferences, the progress in bridging resolution is much slower. The main issue for most researchers aiming to ap- ply statistical algorithms to this task is the lack of training data. While coreference resolution has about 35,000 coref- erent pairs in their standard benchmark dataset OntoNotes (taking into account the transitivity of coreference chains), most datasets for bridging commonly comprise around 400 - 600 pairs (of course, bridging anaphors are also much rarer than coreference anaphors). Note that a benchmark dataset for bridging has not yet been established. In order to tackle the lack of available training data, several smaller corpora could be combined to create a larger corpus re- source, including the corpus presented in this paper. The ISNotes corpus (Markert et al., 2012) contains bridging an- 1Anaphors are marked in bold face, their antecedents are un- …

Content truncated to 500 words