Multitask Learning-Based Neural Bridging Reference Resolution

Proceedings of the 28th International Conference on Computational Linguistics, pages 3534–3546 Barcelona, Spain (Online), December 8-13, 2020 3534 Multitask Learning-Based Neural Bridging Reference Resolution Juntao Yu and Massimo Poesio Queen Mary University of London {juntao.yu, m.poesio}@qmul.ac.uk Abstract We propose a multi task learning-based neural model for resolving bridging references tackling two key challenges. The first challenge is the lack of large corpora annotated with bridging ref- erences. To address this, we use multi-task learning to help bridging reference resolution with coreference resolution. We show that substantial improvements of up to 8 p.p. can be achieved on full bridging resolution with this architecture. The second challenge is the different definitions of bridging used in different corpora, meaning that hand-coded systems or systems using special features designed for one corpus do not work well with other corpora. Our neural model only uses a small number of corpus independent features, thus can be applied to different corpora. Evaluations with very different bridging corpora (ARRAU, ISNOTES, BASHI and SCICORP) sug- gest that our architecture works equally well on all corpora, and achieves the SoTA results on full bridging resolution for all corpora, outperforming the best reported results by up to 36.3 p.p..1 1 Introduction Anaphora resolution (Karttunen, 1976; Webber, 1979; Kamp and Reyle, 1993; Garnham, 2001; Poe- sio et al., 2016) is the aspect of language interpretation concerned with linking nominal expressions to entities in the context of interpretation (or discourse model). As illustrated by (1) (adapted from (Hou et al., 2018)), nominal expressions can be linked to the context in several ways: corefererence (linking [The Bakersfield Supermarket], [The business], [its]), bridging or associative reference (linking [the customers] to [the supermarket]) (Clark, 1975; Prince, 1981; Poesio et al., 2004; Hou et al., 2018), and discourse deixis (linking [the murder] to the event of murdering) (Webber, 1991; Kolhatkar et al., 2018). (1) [The Bakersfield Supermarket] went bankrupt last May. [The business] closed when [[its] old owner] was murdered by [robbers]. [The murder] saddened [the customers]. Bridging reference resolution is the sub-task of anaphora resolution concerned with identifying and resolving bridging references, i.e., anaphoric references to non-identical associated antecedents. Bridg- ing resolution is much less studied than the closely related sub-task of coreference resolution, which has received a lot of attention ((Pradhan et al., 2012; Wiseman et al., 2015; Lee et al., 2017; Lee et al., 2018), to mention just a few recent proposals). One reason for this is the lack of training data. Several corpora have been annotated with bridging reference, including e.g. GNOME (Poesio, 2004), ISNOTES (Markert et al., 2012), SCICORP (R¨osiger, 2016) and BASHI (R¨osiger, 2018), but they are all rather small, with at most around 1k examples of bridging reference. ARRAU (Poesio and Artstein, 2008; Uryupina et al., 2019) is much larger, but still contains only 5.5k bridging pairs. It is challenging to train a learning based system on that amount of data, particularly the new neural models. As a result, the current SoTA systems for full bridging resolution are still rule-based, …

Content truncated to 500 words