Efficient Inference and Structured Learning for Semantic Role Labeling

Efficient Inference and Structured Learning for Semantic Role Labeling Oscar Täckström Google New York oscart@google.com Kuzman Ganchev Google New York kuzman@google.com Dipanjan Das Google New York dipanjand@google.com Abstract We present a dynamic programming algorithm for efficient constrained inference in semantic role labeling. The algorithm tractably captures a majority of the structural constraints exam- ined by prior work in this area, which has re- sorted to either approximate methods or off-the- shelf integer linear programming solvers. In ad- dition, it allows training a globally-normalized log-linear model with respect to constrained conditional likelihood. We show that the dy- namic program is several times faster than an off-the-shelf integer linear programming solver, while reaching the same solution. Furthermore, we show that our structured model results in significant improvements over its local counter- part, achieving state-of-the-art results on both PropBank- and FrameNet-annotated corpora. 1 Introduction Semantic role labeling (henceforth, SRL) is the task of identifying the semantic arguments of predicates in natural language text. Pioneered by Gildea and Jurafsky (2002), this task has been widely investi- gated by the NLP community. There have been two shared tasks at CoNLL 2004 and 2005 focusing on this problem, using PropBank conventions to identify the phrasal arguments of verbal predicates (Palmer et al., 2005; Carreras and Màrquez, 2004, 2005). Since then, there has been work on SRL for nominal pred- icates (Meyers et al., 2004; Gerber and Chai, 2010) and variants that investigated the prediction of se- mantic dependencies rather than phrasal arguments (Surdeanu et al., 2008; Hajiˇc et al., 2009). Here, we present an inference method for SRL, addressing the problem of phrasal argument structure prediction (as opposed to semantic dependencies). In contrast to most prior semantic role labeling work focusing on PropBank conventions, barring notable exceptions such as Meza-Ruiz and Riedel (2009), our framework first performs frame identification, the subtask of disambiguating the predicate frame; this makes our analysis more interpretable. The focus of this paper, however, is the subtask of semantic role labeling, wherein we take a set of (potentially over- lapping) candidate sentential phrases and identify and label them with the semantic roles associated with the predicted frame. This treatment is commonly used in frame semantic parsing (Das et al., 2014; Hermann et al., 2014) and our two-stage framework is able to model both PropBank and FrameNet conventions. Previous work focusing on semantic role labeling imposed several structural constraints warranted by the annotation conventions of the task and other lin- guistic considerations, such as avoiding overlapping arguments and repeated core roles in the final predic- tion. Such global inference often leads to improved results and more meaningful predictions compared to local unconstrained methods (Màrquez et al., 2008). A popular framework for imposing these constraints has been integer linear programming (ILP), wherein the inference problem is specified declaratively (Pun- yakanok et al., 2008). However, ILP-based inference methods often rely on generic off-the-shelf solvers that fail to exploit problem-specific structure (Martins et al., 2011). Instead, we present a dynamic program (DP) that exactly enforces most …

Content truncated to 500 words