Abstract
Fine-grained opinion mining involves identifying opinion boundaries and their associated attributes such as polarity and intensity. Traditional cascaded systems often suffer from error propagation between these sequential steps. This paper introduces a hierarchical sequential learning framework using Conditional Random Fields (CRFs) to jointly model these tasks. By employing a hierarchical parameter-sharing mechanism, the model effectively captures correlations between attributes and reduces the parameter space. Evaluation on the MPQA corpus shows that the hierarchical joint model significantly outperforms traditional cascaded approaches and standard joint models.
Introduction
Fine-grained opinion analysis is a complex multi-step challenge in Natural Language Processing (NLP). It requires identifying the exact boundaries of an opinion expression, determining its polarity (positive, negative, or neutral), and assessing its intensity (high, medium, or low).
Traditionally, these tasks were handled as a "cascade" of separate steps. However, this approach is prone to error propagation: a failure in boundary identification inevitably leads to incorrect attribute assignment. Furthermore, attributes like intensity and polarity are inherently related; the linguistic markers for a "strong" opinion often overlap with those for a specific sentiment. Treating these tasks in isolation ignores valuable correlations.
Methods
The authors propose a Hierarchical Sequential Learning framework based on Conditional Random Fields (CRFs). The core innovation is hierarchical parameter sharing. Instead of treating every combination of polarity and intensity as an independent label, the model organizes them into a tree structure.
In this structure, the top level represents the "Opinion Expression," followed by Polarity labels, and finally Intensity labels. This hierarchy allows the model to learn "Positive" features from all positive examples regardless of intensity, making it more robust, especially for rare label combinations. This approach also significantly reduces the number of parameters compared to a flat joint model (from 110 to 46 sets), acting as a form of regularization. The model incorporates standard NLP features like POS tags and WordNet hypernyms, alongside expression-level heuristics to smooth token-level data.
Results
The study utilized the Multi-Perspective Question Answering (MPQA) corpus, a benchmark containing news articles annotated for subjective expressions. The hierarchical model was compared against two baselines: a cascaded model (Baseline-1) and a joint model without hierarchy (Baseline-2).
As shown in the overall extraction results, the hierarchical model achieved an F-measure of 62.0, outperforming the cascaded baseline (58.9) and the simple joint model (60.5). These results confirm that joint learning is superior to the cascaded approach and that the hierarchy provides additional performance gains.
Specific attribute analysis revealed that the model performed best on "Negative" opinions (reaching an F-measure of 52.8), likely due to the higher prevalence of negative examples in the dataset. Conversely, intensity remains a more difficult attribute to classify; the authors found that distinguishing between "Medium" and "Low" intensity is often a subtle, subjective task even for human annotators.
Conclusion
The hierarchical joint model effectively addresses the limitations of cascaded systems by reducing error propagation and leveraging attribute correlations. While intensity classification remains a stubborn challenge, the hierarchical parameter-sharing logic provides a robust framework for multi-task NLP problems. Future work may involve exploring more complex hierarchies or integrating additional attributes, such as the target of the opinion expression.
Summary updated