Predicate Argument Alignment using a Global Coherence Model

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 11–20, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Predicate Argument Alignment using a Global Coherence Model Travis Wolfe, Mark Dredze, and Benjamin Van Durme Johns Hopkins University Baltimore, MD, USA Abstract We present a joint model for predicate argu- ment alignment. We leverage multiple sources of semantic information, including temporal ordering constraints between events. These are combined in a max-margin framework to find a globally consistent view of entities and events across multiple documents, which leads to improvements over a very strong local baseline. 1 Introduction Natural language understanding (NLU) requires analysis beyond the sentence-level. For example, an entity may be mentioned multiple times in a dis- course, participating in various events, where each event may itself be referenced elsewhere in the text. Traditionally the task of coreference resolution has been defined as finding those entity mentions within a single document that co-refer, while cross- document coreference resolution considers a wider discourse context across many documents, yet still pertains strictly to entities. Predicate argument alignment, or entity-event cross-document coreference resolution, enlarges the set of possible co-referent elements to include the mentions of situations in which entities participate. This expanded definition drives practitioners to- wards a more complete model of NLU, where sys- tems must not only consider who is mentioned, but also what happened. However, despite the drive to- wards an expanded notion of discourse, models typ- ically are formulated with strong notions of local- independence: viewing a multi-document task as one limited to individual pairs of sentences. This creates a mis-match between the goals of such work – considering entire documents – with the systems – consider individual sentences. In this work, we consider a system that takes a document level view in considering coreference for entities and predictions: the task of predicate ar- gument linking. We treat this task as a global in- ference problem, leveraging multiple sources of se- mantic information identified at the document level. Global inference for this problem is mostly unex- plored, with the exception of Lee et al. (2012) (dis- cussed in § 8). Especially novel here is the use of document-level temporal constraints on events, rep- resenting a next step forward on the path to full un- derstanding. Our approach avoids the pitfalls of local infer- ence while still remaining fast and exact. We use the pairwise features of a very strong predicate argu- ment aligner (Wolfe et al., 2013) (competitive with the state-of-the-art (Roth, 2014)), and add quadratic factors that constrain local decisions based on global document information. These global factors lead to superior performance compared to the previous state-of-the-art. We release both our code and data.1 2 Model Consider the two sentences from the document pair shown in Figure 1. These sentences describe the same event, although with different details. The source sentence has four predicates and four ar- guments, while the target has three predicates and three …

Content truncated to 500 words