A Joint Model for Extended Semantic Role Labeling

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 129–139, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics A Joint Model for Extended Semantic Role Labeling Vivek Srikumar and Dan Roth University of Illinois, Urbana-Champaign Urbana, IL 61801 {vsrikum2, danr}@illinois.edu Abstract This paper presents a model that extends se- mantic role labeling. Existing approaches in- dependently analyze relations expressed by verb predicates or those expressed as nominal- izations. However, sentences express relations via other linguistic phenomena as well. Fur- thermore, these phenomena interact with each other, thus restricting the structures they artic- ulate. In this paper, we use this intuition to define a joint inference model that captures the inter-dependencies between verb seman- tic role labeling and relations expressed us- ing prepositions. The scarcity of jointly la- beled data presents a crucial technical chal- lenge for learning a joint model. The key strength of our model is that we use existing structure predictors as black boxes. By en- forcing consistency constraints between their predictions, we show improvements in the per- formance of both tasks without retraining the individual models. 1 Introduction The identification of semantic relations between sentence constituents has been an important task in NLP research. It finds applications in various natural language understanding tasks that require complex inference going beyond the surface representation. In the literature, semantic role extraction has been studied mostly in the context of verb predicates, us- ing the Propbank annotation of Palmer et al. (2005), and also for nominal predicates, using the Nombank corpus of Meyers et al. (2004). However, sentences express semantic relations through other linguistic phenomena. For example, consider the following sentence: (1) The field goal by Brien changed the game in the fourth quarter. Verb centered semantic role labeling would identify the arguments of the predicate change as (a) The field goal by Brien (A0, the causer of the change), (b) the game (A1, the thing changing), and (c) in the fourth quarter (temporal modifier). However, this does not tell us that the scorer of the field goal was Brien, which is expressed by the preposition by. Also, note that the in indicates a temporal relation, which overlaps with the verb’s analysis. In this paper, we propose an extension of the stan- dard semantic role labeling task to include relations expressed by lexical items other than verbs and nom- inalizations. Further, we argue that there are interac- tions between the different phenomena which sug- gest that there is a benefit in studying them together. However, one key challenge is that large jointly la- beled corpora do not exist. This motivates the need for novel learning and inference schemes that ad- dress the data problem and can still benefit from the interactions among the phenomena. This paper has two main contributions. 1. From the machine learning standpoint, we pro- pose a joint inference scheme to combine exist- ing structure predictors for multiple linguistic phenomena. We do so using hard constraints that involve only the labels …

Content truncated to 500 words