Context-aware Frame-Semantic Role Labeling

Context-aware Frame-Semantic Role Labeling Michael Roth and Mirella Lapata School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB {mroth,mlap}@inf.ed.ac.uk Abstract Frame semantic representations have been useful in several applications ranging from text-to-scene generation, to question answer- ing and social network analysis. Predicting such representations from raw text is, how- ever, a challenging task and corresponding models are typically only trained on a small set of sentence-level annotations. In this pa- per, we present a semantic role labeling sys- tem that takes into account sentence and dis- course context. We introduce several new fea- tures which we motivate based on linguistic insights and experimentally demonstrate that they lead to significant improvements over the current state-of-the-art in FrameNet-based se- mantic role labeling. 1 Introduction The goal of semantic role labeling (SRL) is to iden- tify and label the arguments of semantic predicates in a sentence according to a set of predefined re- lations (e.g., “who” did “what” to “whom”). In addition to providing definitions and examples of role labeled text, resources like FrameNet (Ruppen- hofer et al., 2010) group semantic predicates into so- called frames, i.e., conceptual structures describing the background knowledge necessary to understand a situation, event or entity as a whole as well as the roles participating in it. Accordingly, semantic roles are defined on a per-frame basis and are shared among predicates. In recent years, frame representations have been successfully applied in a range of downstream tasks, including question answering (Shen and Lapata, 2007), text-to-scene generation (Coyne et al., 2012), stock price prediction (Xie et al., 2013), and so- cial network extraction (Agarwal et al., 2014). Whereas some tasks directly utilize information encoded in the FrameNet resource, others make use of FrameNet indirectly through the output of SRL systems that are trained on data annotated with frame-semantic representations. While ad- vances in machine learning have recently given rise to increasingly powerful SRL systems follow- ing the FrameNet paradigm (Hermann et al., 2014; T¨ackstr¨om et al., 2015), little effort has been devoted to improve such models from a linguistic perspec- tive. In this paper, we explore insights from the lin- guistic literature suggesting a connection between discourse and role labeling decisions and show how to incorporate these in an SRL system. Although early theoretical work (Fillmore, 1976) has recog- nized the importance of discourse context for the assignment of semantic roles, most computational approaches have shied away from such considera- tions. To see how context can be useful, consider as an example the DELIVERY frame, which states that a THEME can be handed off to either a RECIPIENT or “more indirectly” to a GOAL. While the distinc- tion between the latter two roles might be clear for some fillers (e.g., people vs. locations), there are oth- ers where both roles are equally plausible and addi- tional information is required to resolve the ambigu- ity (e.g., countries). If we hear about a letter being delivered to Greece, for instance, reliable cues might be whether the sender is a …

Content truncated to 500 words