Generative models for semantic role labeling

Generative Models for Semantic Role Labeling Cynthia A. Thompson cindi@cs.utah.edu Siddharth Patwardhan sidd@cs.utah.edu School of Computing University of Utah Salt Lake City, UT 84112 Carolin Arnold csarnold@cs.utah.edu Abstract This paper describes the four entries from the Uni- versity of Utah in the semantic role labeling task of SENSEVAL-3. All the entries took a statisti- cal machine learning approach, using the subset of the FrameNet corpus provided by SENSEVAL-3 as training data. Our approach was to develop a model of natural language generation from seman- tics, and train the model using maximum likelihood and smoothing. Our models performed satisfacto- rily in the competition, and can flexibly handle vary- ing permutations of provided versus inferred infor- mation. 1 Introduction The goal in the SENSEVAL-3 semantic role labeling task is to identify roles and optionally constituent boundaries for each role, given a natural language sentence, target, and frame. The Utah approach to this task is to apply machine learning techniques to create a model capable of semantically analyz- ing unseen sentences. We have developed a set of generative models (Jordan, 1999) that have the ad- vantages of flexibility, power, and ease of applica- bility for semi-supervised learning scenarios. We can supplement any of the generative models with a constituent classifier that determines, given a sen- tence and parse, which parse constituents are most likely to correspond to a role. We apply the com- bination to the “hard,” or restricted version of the role labeling task, in which the system is provided only with the sentence, target, and frame, and must determine which sentence constituents to label with roles. We discuss our overall model, the constituent classifier we use in the hard task, and the classifier’s use at role-labeling time. We entered four sets of answers, as discussed in Section 5. The first two correspond to the “easy” task, in which the role- bearing constituents – those parts of the sentence corresponding to a role – are provided to the system with the target and frame. The second two are vari- ants for the “hard” task. Finally, we discuss Future Work and conclude the paper. 2 Role Labeler Our general approach is to use a generative model defining a joint probability distribution over targets, frames, roles, and constituents. The advantage of such a model is its generality: it can determine the probability of any subset of the variables given val- ues for the others. Three of our entries used the generative model illustrated in Figure 1, and the fourth used a model grouping all roles together, as described further below. The first model functions T F R1 R2 · · · Rn C1 C2 · · · Cn Figure 1: First Order Model. as follows. First, a target, T, is chosen, which then generates a frame, F. The frame generates a (lin- earized) role sequence, R1 through Rn which in turn generates each constituent of the sentence, C1 through Cn. Note that, conditioned on a particular frame, the model is just a first-order Hidden …

Content truncated to 500 words