[Paper Review] An MRC Framework for Semantic Role Labeling

Nan Wang, Jiwei Li+6 more|arXiv (Cornell University)|Sep 14, 2021

Topic Modeling55 references4 citations

TL;DR

This paper proposes a machine reading comprehension (MRC) framework for semantic role labeling (SRL) that unifies predicate disambiguation and argument labeling by modeling them as multiple-choice and extractive MRC tasks, respectively, using semantic descriptions from frame files. The approach achieves state-of-the-art or competitive results on CoNLL2005, CoNLL2009, and CoNLL2012 benchmarks while incorporating role semantics and improving computational efficiency via a role prediction module.

ABSTRACT

Semantic Role Labeling (SRL) aims at recognizing the predicate-argument structure of a sentence and can be decomposed into two subtasks: predicate disambiguation and argument labeling. Prior work deals with these two tasks independently, which ignores the semantic connection between the two tasks. In this paper, we propose to use the machine reading comprehension (MRC) framework to bridge this gap. We formalize predicate disambiguation as multiple-choice machine reading comprehension, where the descriptions of candidate senses of a given predicate are used as options to select the correct sense. The chosen predicate sense is then used to determine the semantic roles for that predicate, and these semantic roles are used to construct the query for another MRC model for argument labeling. In this way, we are able to leverage both the predicate semantics and the semantic role semantics for argument labeling. We also propose to select a subset of all the possible semantic roles for computational efficiency. Experiments show that the proposed framework achieves state-of-the-art or comparable results to previous work. Code is available at \url{https://github.com/ShannonAI/MRC-SRL}.

Motivation & Objective

• To address the lack of semantic integration between predicate disambiguation and argument labeling in traditional SRL approaches.
• To leverage semantic information from frame files—specifically, the meanings of predicate senses and argument roles—within a unified framework.
• To improve model robustness in low-resource and out-of-domain settings by incorporating label semantics.
• To reduce computational overhead in argument labeling by filtering irrelevant semantic roles through a role prediction module.

Proposed method

• Formalize predicate disambiguation as multiple-choice MRC, where candidate predicate senses (with their descriptions) are treated as answer options.
• Use the predicted predicate sense to determine relevant semantic roles for the target predicate.
• Transform argument labeling into extractive MRC by constructing queries from semantic role descriptions (e.g., 'causer of motion') to extract argument spans.
• Introduce a role prediction module that selects a subset of the most probable semantic roles (e.g., top λN roles) to reduce computation.
• Use RoBERTa-Base as the base model for all MRC components, with fine-tuning on CoNLL benchmarks.
• Apply filtering using lemmatized predicates and learned role probabilities to avoid encoding all possible <label, sentence> pairs.

Experimental results

Research questions

• RQ1 Can modeling predicate disambiguation and argument labeling as MRC tasks improve semantic coherence between the two subtasks?
• RQ2 Does incorporating explicit semantic descriptions of predicate senses and argument roles enhance performance, especially in low-resource or out-of-domain settings?
• RQ3 How effective is a role prediction module in reducing computational cost without sacrificing labeling accuracy?
• RQ4 What is the impact of role semantics on argument labeling performance when training and testing data differ in domain?

Key findings

• The proposed MRC framework achieves state-of-the-art or competitive results on CoNLL2005, CoNLL2009, and CoNLL2012 benchmarks.
• Including role semantics in argument labeling improves F1 scores by 0.6 points on the development set (88.5 vs. 88.2) and 0.6 points on the out-of-domain Brown test set (83.8 vs. 83.2).
• When role prediction recall is increased from 0.90 to 0.99, argument labeling F1 improves from 87.3 to 88.6, indicating reduced error propagation.
• In low-resource scenarios, the MRC model with role semantics outperforms a baseline BIO tagging model without semantic priors, especially at low data proportions.
• The role prediction module achieves 4x to 5.1x speedup in computation by filtering irrelevant roles, with λ values of 5, 4.2, and 5.5 on CoNLL2005, CoNLL2009, and CoNLL2012, respectively.