Enhance Weakly-Supervised Aspect Detection with External Knowledge (Student Abstract)

Zhuoming Zheng, Yi Cai, Liuwu LiAAAI 2022 · DOI

Introduction

Aspect detection—the task of identifying specific product features discussed in user reviews—is a cornerstone of sentiment analysis and opinion mining. Traditional seed-driven methods rely on human-provided keywords to categorize sentences. However, these methods often struggle with two primary issues: the "long tail" of infrequent technical terms (e.g., "AHCI") and "Miscellaneous" segments that contain no specific aspect information. This paper proposes a weakly-supervised framework that leverages external common-sense knowledge to bridge these semantic gaps.

Methodology

The core of the proposed approach is the integration of structured external knowledge from ConceptNet to enrich the representation of review segments.

Knowledge-Enhanced Graph Construction

The model identifies entities within review segments and maps them to nodes in ConceptNet. By exploring relationships such as "part of" or "related to," the system can link rare terms to broader aspect categories like "Hardware." For instance, while "touchscreen" might be rare in a specific dataset, its permanent relationship to "hardware" in a knowledge graph allows the model to categorize it correctly without extensive training data.

Figure 1: Conceptual overview of linking review segments to aspect categories via ConceptNet.

Model Architecture

The framework utilizes a Relational Graph Convolutional Network (R-GCN) to generate latent representations of entities and their relationships, employing DistMult factorization to score graph triples. To refine these representations, the authors use contrastive learning, which maximizes the similarity between a reconstructed segment and its original form while distancing it from unrelated segments.

Knowledge Distillation

To combine structured knowledge with deep semantic understanding, a Teacher-Student distillation process is employed. A complex "Teacher" model (incorporating the Knowledge Graph) trains a "Student" BERT model. This ensures the final system retains the benefits of external knowledge while leveraging the powerful pre-trained language modeling capabilities of BERT.

Experimental Evaluation

The researchers evaluated the model using the OpoSum dataset, focusing on the "TVs" and "Keyboards" domains to test performance on specialized vocabulary.

Performance Comparison

The proposed model, Ours(Stu.), was compared against several strong baselines, including BERT (2019) and AspMem (2020). The results demonstrate that the inclusion of external knowledge significantly improves detection accuracy, particularly in domains with technical jargon.

Figure 3: Performance comparison (Micro-F1) against state-of-the-art models on OpoSum datasets.

Ablation studies further confirmed the value of each component. Adding the Knowledge Graph (KG) to a base model improved the average Micro-F1 from 61.7 to 62.9, while the final distillation into BERT achieved a peak score of 64.8.

Discussion and Conclusion

The study highlights that external knowledge acts as a vital bridge for identifying "implicit" aspects. Furthermore, the model effectively handles the "Miscellaneous" problem by identifying segments that lack strong connections to domain-specific knowledge sub-graphs.

While the current implementation is limited to "one-hop" entity extraction from ConceptNet, future work could explore multi-hop relationships to capture deeper context. Overall, this research demonstrates that structured common-sense knowledge and distributional semantics are highly complementary for fine-grained text analysis tasks.

Summary updated

BibTeX
@inproceedings{zheng2022enhance,
  author = {Zhuoming Zheng and Yi Cai and Liuwu Li},
  title = {{Enhance Weakly-Supervised Aspect Detection with External Knowledge (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21696},
}
Figure