Enhancing Opinion Role Labeling with Semantic-Aware Word Representations from Semantic Role Labeling

Meishan Zhang, Peili Liang, Guohong Fu

Abstract

Opinion Role Labeling (ORL) is a critical task in fine-grained sentiment analysis, aimed at identifying the holder, target, and expression of an opinion. While linguistically related to Semantic Role Labeling (SRL), direct integration of SRL outputs into ORL systems often suffers from error propagation and role mismatches. This paper introduces SRL-SAWR, a method that extracts continuous, semantic-aware word representations from a neural SRL encoder. By fine-tuning these representations for the ORL task, the model achieves a 5.4% absolute improvement in F1-score over a strong baseline on the MPQA 2.0 dataset.

Introduction

Fine-grained opinion mining requires identifying three core components: the holder (the source of the opinion), the target (the entity being evaluated), and the expression (the lexical indicator of the opinion). This task, known as Opinion Role Labeling (ORL), shares structural similarities with Semantic Role Labeling (SRL), which identifies the semantic arguments (agents, patients) of a predicate.

Despite these similarities, two primary challenges hinder the effective use of SRL for ORL:

  1. Linguistic Mismatches: Semantic roles do not always map one-to-one to opinion roles. For instance, in the sentence "The report embarrassed the White House," the "White House" is a semantic patient (ARG1) but serves as the opinion holder.
  2. Error Propagation: Traditional pipeline approaches use discrete SRL tags as features for ORL. Any errors in the initial SRL stage are directly inherited by the ORL model, leading to degraded performance.
Distribution of semantic roles for opinion holders and targets, illustrating the complex mapping between SRL and ORL.

Methodology: SRL-SAWR

The proposed SRL-SAWR (Semantic-Aware Word Representations) model addresses these issues by integrating SRL information at the representation level rather than the output level.

Architecture

The core of the system is a neural architecture that leverages a pre-trained, state-of-the-art SRL encoder. Instead of utilizing the final discrete SRL tags, the model extracts the internal hidden vector representations from the SRL encoder. These vectors capture rich, "soft" semantic information that is often lost in hard classification.

The SRL-SAWR architecture, showing how hidden representations from an SRL encoder are integrated into the ORL sequence labeling pipeline.

These semantic-aware vectors are concatenated with standard word embeddings and passed through a projection matrix to serve as input for the ORL model—a deep Bi-LSTM followed by a Conditional Random Field (CRF) layer. Crucially, the parameters of the SRL encoder are fine-tuned during the training of the ORL model, allowing the representations to adapt to the specific nuances of opinion mining.

Experimental Setup

The model was evaluated on the MPQA 2.0 corpus, a standard benchmark for fine-grained opinion mining containing 482 documents. The auxiliary SRL encoder was pre-trained on the CoNLL-2012 shared task dataset, which provides over 250,000 annotated predicates from OntoNotes v5.0.

Results and Discussion

The SRL-SAWR model demonstrated significant performance gains across all metrics compared to a standard Bi-LSTM baseline and a multi-task learning (FS-MTL) approach.

Model Exact F1 (Overall) Holder F1 Target F1
Baseline 58.30 73.07 42.70
FS-MTL 61.51 75.58 46.40
SRL-SAWR (Proposed) 63.74 76.95 50.50

The most notable improvement was observed in target identification, which saw nearly an 8% absolute increase in F1-score. This suggests that the semantic-aware representations are particularly effective at capturing the complex relationships between opinion expressions and their targets.

Key Insights

  • Representations over Labels: Continuous vectors retain ambiguous semantic information that helps the ORL layer resolve difficult cases where a hard SRL tag might be misleading.
  • Handling Inconsistency: The neural mapping learned through fine-tuning handles "inconsistent" cases—where SRL and ORL roles diverge—more robustly than heuristic or multi-task methods.
  • Reduced Sensitivity to Errors: By avoiding a pipeline of discrete labels, the model is less prone to the "all-or-nothing" failures typical of systems that rely on pre-computed SRL tags.

Conclusion

This work demonstrates that leveraging internal representations from related semantic tasks can significantly boost performance in opinion mining. While the current approach introduces some computational overhead due to the dual-encoder architecture, the gains in accuracy justify the cost. Future research could explore similar representation-sharing techniques using other auxiliary tasks like dependency parsing or cross-lingual transfer learning for low-resource scenarios.

Summary updated

BibTeX
@inproceedings{zhang2019enhancing,
  author = {Meishan Zhang and Peili Liang and Guohong Fu},
  title = {{Enhancing Opinion Role Labeling with Semantic-Aware Word Representations from Semantic Role Labeling}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
  doi = {10.18653/v1/n19-1066},
}
Figure