Abstract
Aspect-level sentiment analysis is a fine-grained sentiment analysis task designed to identify the sentiment polarity of specific targets in a sentence. However, this task is rarely used in drug reviews. Some models for this task ignore the impact of target semantics, and others do not perform well because the datasets are relatively smaller. Therefore, we propose a Pretraining and Multi-task learning model based on Double BiGRU (PM-DBiGRU). In PM-DBiGRU, we first use the pretrained weight learned from short text-level drug review sentiment classification task to initialize related weight of our model. Then two BiGRU networks are applied to generate the bidirectional semantic representations of the target and drug review, and attention mechanism is used to obtain the target-specific representation for aspect-level drug review. The multi-task learning is further utilized to transfer the helpful domain knowledge from the short text-level drug review corpus. We also propose a dataset SentiDrugs for aspect-level drug review sentiment classification, in which each review may contain one or more targets. Experimental results on SentiDrugs demonstrate that our approach can improve the performance of aspect-level drug reviews sentiment classification compared with other state-of-the-art architectures.
Introduction
Sentiment analysis typically evaluates the overall sentiment of a sentence. However, in specialized domains like medicine, a single sentence often contains multiple targets with differing sentiment polarities. Aspect-Level Sentiment Analysis (ALSA) addresses this by identifying sentiment toward specific targets. For instance, in the statement "This drug works great for water retention, but its side effect is severe," the sentiment is positive for "water retention" but negative for "side effect."
Applying ALSA to medical drug reviews presents unique challenges. First, manually annotating medical reviews is resource-intensive, leading to data scarcity. Second, medical targets are often multi-word phrases (e.g., "cholesterol and blood pressure"), making them harder to model than simple product features. Finally, sentiment is often domain-specific; a word like "low" may be positive for blood sugar but negative for product quality.
Methodology: The PM-DBiGRU Model
To address these challenges, we propose PM-DBiGRU (Pretraining and Multi-task learning based on Double BiGRU). The architecture incorporates several key innovations:
- Double BiGRU Architecture: Unlike standard models that use a single BiGRU for the entire sentence, PM-DBiGRU employs two. One BiGRU processes the sentence context, while the second specifically models the semantics of the target aspect. This ensures the model captures the specific meaning of complex medical targets.
- Knowledge Transfer via Pretraining: To mitigate the small dataset issue, the model is initialized with weights pretrained on a large-scale general drug review dataset. This allows the model to learn domain-specific language patterns before fine-tuning on the aspect-level task.
- Multi-task Learning: The model simultaneously learns general sentiment classification and fine-grained aspect-level classification. This parameter sharing facilitates better generalization by leveraging broader domain knowledge.
- Attention Mechanism: An attention layer dynamically links the target representation to the sentence context, allowing the model to focus on sentiment-bearing words (e.g., verbs like "receded" or adjectives like "severe") relevant to the specific target.
Datasets and Experimental Setup
The research utilizes two primary data sources. The first is SentiDrugs, a newly introduced dataset containing 4,028 manually annotated target-sentiment pairs derived from 4,200 drug reviews. The second is Drugs.com, a large public dataset used for pretraining and multi-task learning components to provide broader domain knowledge.
As shown in the figures above, the SentiDrugs dataset features a variety of target lengths and review lengths, highlighting the complexity of the ALSA task in this domain. Most targets are short, but a significant portion consists of multi-word phrases that require robust semantic modeling.
Key Results and Discussion
PM-DBiGRU was evaluated against several state-of-the-art baselines, including IAN, MemNet, and TNet.
- Performance: PM-DBiGRU achieved an accuracy of 78.26% and a Macro-F1 score of 77.75%, outperforming the best baseline (PRET+MULT) which reached 76.79% accuracy.
- Ablation Study: Removing pretraining or multi-task learning led to significant performance declines, confirming that knowledge transfer is critical for success on small medical datasets.
- Semantic Importance: The use of a dedicated BiGRU for targets proved superior to simple word vector averaging, particularly for multi-word medical terms. The attention mechanism successfully identified that sentiment in medical reviews is often carried by verbs (e.g., "pain receded") rather than just adjectives.
Conclusion and Future Work
While PM-DBiGRU shows significant promise, challenges remain in handling implicit sentiment and reviews requiring deep background medical knowledge. Future work will focus on integrating external medical knowledge bases and improving the model's ability to handle complex syntactic structures in long, rambling reviews.
Summary updated