Bridging resolution: Task definition, corpus resources and rule-based experiments

Proceedings of the 27th International Conference on Computational Linguistics, pages 3516–3528 Santa Fe, New Mexico, USA, August 20-26, 2018. 3516 Bridging resolution: Task definition, corpus resources and rule-based experiments Ina R¨osiger, Arndt Riester and Jonas Kuhn Institute for Natural Language Processing University of Stuttgart, Germany {roesigia,arndt,jonas}@ims.uni-stuttgart.de Abstract Recent work on bridging resolution has so far been based on the corpus ISNotes (Markert et al., 2012), as this was the only corpus available with unrestricted bridging annotation. Hou et al.’s (2014) rule-based system currently achieves state-of-the-art performance on this corpus, as learning-based approaches suffer from the lack of available training data. Recently, a number of new corpora with bridging annotations have become available. To test the generalisability of the approach by Hou et al. (2014), we apply a slightly extended rule-based system to these corpora. Besides the expected out-of-domain effects, we also observe low performance on some of the in- domain corpora. Our analysis shows that this is the result of two very different phenomena being defined as bridging, which we call referential and lexical bridging. We also report that filtering out gold or predicted coreferent anaphors before applying the bridging resolution system helps improve bridging resolution. 1 Introduction Bridging is an anaphoric phenomenon where the interpretation of a bridging anaphor, sometimes also called associative anaphor (Hawkins, 1978), is based on the non-identical associated antecedent. The corresponding NLP task of bridging resolution is about linking these anaphoric noun phrases and their antecedents, which 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) Our correspondent in Egypt is reporting that the opposition is holding a rally against the consti- tutional referendum.1 One can think of bridging anaphors as expressions with an implicit argument, e.g. the opposition (in Egypt). Compared to coreference resolution, which has become one of the standard NLP tasks, the progress in bridging resolution is much slower. The main issue for most researchers aiming to apply statistical algorithms to this task is the lack of training data, as well as the rather diverse bridging defi- nitions and annotations. The resolution of bridging links is important because it can prove beneficial in tasks which use the concept of textual coherence, for example Barzilay and Lapata’s (2008) entity grid or Hearst’s (1994) text segmentation. Note that while a benchmark dataset for bridging has not yet been established, most recent work is based on the ISNotes corpus (Markert et al., 2012), which contains Wall Street Journal articles. Full bridging resolution on this corpus has been investigated in Hou et al. (2014), following earlier experi- ments on the subtasks of bridging anaphor detection (Hou et al., 2013a) and antecedent selection (Hou et al., 2013b). Apart from this, there is some work on bridging detection as a subclass of information status classification, where bridging is typically a category with low annotator agreement and low detection ac- curacy (Markert et al., 2012; Rahman and Ng, 2012; …

Content truncated to 500 words