A Joint Model for Extended Semantic Role Labeling

Vivek Srikumar, Dan RothEMNLP 2011 · OpenReview

Abstract

This paper presents a model that extends semantic role labeling. Existing approaches independently analyze relations expressed by verb predicates or those expressed as nominalizations. However, sentences express relations via other linguistic phenomena as well. Furthermore, these phenomena interact with each other, thus restricting the structures they articulate. In this paper, we use this intuition to define a joint inference model that captures the inter-dependencies between verb semantic role labeling and relations expressed using prepositions. The scarcity of jointly labeled data presents a crucial technical challenge for learning a joint model. The key strength of our model is that we use existing structure predictors as black boxes. By enforcing consistency constraints between their predictions, we show improvements in the performance of both tasks without retraining the individual models.

Introduction

Semantic Role Labeling (SRL) is a fundamental task in natural language processing, aimed at identifying the semantic relationships between predicates and their arguments—essentially determining "who did what to whom." Traditionally, research has focused almost exclusively on relations centered around verb predicates (e.g., Propbank) or nominalizations (e.g., Nombank). However, a significant portion of semantic information in natural language is conveyed through other linguistic markers, most notably prepositions.

For instance, in the phrase "the field goal by Brien," a standard verb-centric system might identify "Brien" as part of an agent, but it fails to explicitly capture that "Brien" is the scorer—a specific relation signaled by the preposition "by." This paper addresses the challenge of extending SRL to include these prepositional relations. The primary obstacles are the scarcity of large-scale corpora labeled for both verb and preposition roles and the complex interdependencies between these linguistic phenomena.

Methodology

The authors propose a Joint Inference Model that leverages Integer Linear Programming (ILP) to enforce consistency between independent "black-box" predictors. This approach is unique because it avoids the need for massive jointly labeled datasets or the retraining of existing state-of-the-art models.

Preposition Role Inventory

To make the task semantically meaningful and computationally tractable, the authors collapsed fine-grained dictionary senses from The Preposition Project into a set of "Preposition Roles." These roles, such as TEMPORAL, LOCATION, and PARTWHOLE, capture the core semantic contribution of prepositions across different contexts.

Preposition roles and their distribution in the training and test sets.

Joint Inference via ILP

The global model integrates a verb SRL system and a preposition classifier. Because these systems may operate on different numeric scales, the authors introduce scaling parameters ($\lambda$) to calibrate their outputs. The core of the joint model is a set of declarative linguistic constraints that encode the interdependencies between verb arguments and prepositional roles.

For example, if a verb SRL system identifies a phrase as a temporal modifier (AM-TMP), the preposition within that phrase must be assigned a TEMPORAL role. These constraints were mined automatically from data and supplemented with hand-written rules based on linguistic intuition to ensure that the combined output is linguistically "legal."

Example of a linguistic constraint mapping verb arguments to preposition roles.

Results

The joint inference model demonstrates performance gains across both tasks by allowing the confidence of one system to correct the errors of another. The F1 score for verb SRL improved from a baseline of 76.22 to 77.07. Preposition role classification accuracy also saw an increase, rising from 67.82% to 68.39% in standard settings.

The most significant impact was observed in domain adaptation scenarios. When a preposition classifier trained on the SemEval dataset was applied to the Penn Treebank, joint inference boosted accuracy from 53.29% to 56.22%. This nearly 3% absolute improvement highlights the robustness of linguistic constraints across different domains, even without additional training data for the underlying classifiers.

Conclusion

This work demonstrates that enforcing linguistic consistency through joint inference is a powerful technique for improving semantic analysis. By treating existing models as black boxes and layering an ILP-based inference step, the authors successfully captured the interdependencies between verbs and prepositions. This framework is highly extensible and provides a path toward a more comprehensive "Extended SRL" that encompasses a wider array of linguistic phenomena, such as compound nouns or possessives.

Summary updated

Abstract

This paper presents a model that extends semantic role labeling. Existing approaches independently analyze relations expressed by verb predicates or those expressed as nominalizations. However, sentences express relations via other linguistic phenomena as well. Furthermore, these phenomena interact with each other, thus restricting the structures they articulate. In this paper, we use this intuition to define a joint inference model that captures the inter-dependencies between verb semantic role labeling and relations expressed using prepositions. The scarcity of jointly labeled data presents a crucial technical challenge for learning a joint model. The key strength of our model is that we use existing structure predictors as black boxes. By enforcing consistency constraints between their predictions, we show improvements in the performance of both tasks without retraining the individual models.
BibTeX
@inproceedings{srikumar2011joint,
  author = {Vivek Srikumar and Dan Roth},
  title = {{A Joint Model for Extended Semantic Role Labeling}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2011},
  url = {https://openreview.net/forum?id=B1b3oMzOWr},
}
Figure