Abstract
We present an approach for the joint extraction of entities and relations in the context of opinion recognition and analysis. We identify two types of opinion-related entities — expressions of opinions and sources of opinions — along with the linking relation that exists between them. Inspired by Roth and Yih (2004), we employ an integer linear programming approach to solve the joint opinion recognition task, and show that global, constraint-based inference can significantly boost the performance of both relation extraction and the extraction of opinion-related entities. Performance further improves when a semantic role labeling system is incorporated. The resulting system achieves F-measures of 79 and 69 for entity and relation extraction, respectively, improving substantially over prior results in the area.
Introduction: The Challenge of Opinion Recognition
In Natural Language Processing (NLP), opinion recognition extends beyond simple sentiment analysis. It requires identifying the "source" (who holds the opinion) and the "expression" (the subjective statement itself), as well as the link between them. Traditionally, these components were extracted using a "pipeline" approach: separate models would identify entities and then a subsequent model would attempt to link them.
However, this sequential method ignores the inherent dependencies between these elements. For instance, a high-confidence opinion expression should logically increase the probability that a nearby noun phrase is its source. Conversely, an isolated source with no linked expression is likely a false positive. This research addresses these shortcomings by proposing a unified system that extracts entities and relations simultaneously through global inference.
Methodology: From Local Guesses to Global Optimization
The proposed system employs a three-phase architecture to move from local candidate generation to a globally consistent solution:
- Entity Extraction: Two separate Conditional Random Fields (CRFs) identify potential opinion expressions and sources. To maximize recall, the system collects the $n$-best sequences rather than just the top result.
- Relation Classification: A Maximum Entropy (MaxEnt) binary classifier evaluates all possible pairs of extracted entities to determine the likelihood of a "link" relation.
- Global Inference via Integer Linear Programming (ILP): This core component formulates the task as an optimization problem. The goal is to maximize the total confidence scores from the previous phases while adhering to logical constraints.
- Hard Constraints: These ensure structural integrity, such as requiring that linked entities must be extracted and preventing overlapping entity predictions.
- Soft Constraints: These incorporate linguistic heuristics, such as the tendency for an opinion expression to have a single source.
Additionally, the authors integrate Semantic Role Labeling (SRL). Given that many opinion-source relations follow predicate-argument structures (e.g., "[Source] claims [Opinion]"), SRL data serves as both a feature for classification and a source of additional constraints for the ILP solver.
Key Results and Performance
The joint extraction approach demonstrated significant improvements over traditional baseline heuristics, such as linking an opinion to its nearest source candidate.
As shown in the comparative data, the ILP framework acts as a powerful filter, discarding incorrect entities that standalone classifiers might retain. By enforcing global consistency, the system achieved the following benchmarks:
- Opinion Expression Extraction: Reached an F-measure of 82.0, comparable to human-level annotation agreement.
- Source Extraction: Achieved an F-measure of 78.1, a substantial improvement over the previous state-of-the-art (69.4).
- Link Relation: Achieved an F-measure of 68.9.
Insights and Future Directions
The success of this model highlights the power of global constraints in NLP. By intentionally "over-generating" candidates in the initial phase and using ILP as a high-precision selector, the system effectively manages the recall-precision tradeoff. Furthermore, the inclusion of SRL provides a significant performance boost by leveraging existing linguistic structures.
Despite these gains, limitations remain. The model currently struggles with "implicit" sources—cases where the source is not explicitly named in the text—as the ILP requires an explicit link to retain an entity. Future work will focus on extending the framework to handle these implicit relations and exploring dependencies that span across multiple sentences.
Summary updated