A Review of Medical Domain Sentiment Analysis: Interpretability, Contextual Aspect‑Sentiment Relations, Noisy Labels, and Domain Lexicon Construction

Introduction Sentiment analysis, a branch of text classification, aims to determine the polarity of subjective texts. In the medical domain, challenges arise because medical entities can serve both as aspects and sentiment words, sentiment dictionaries often label disease terms as negative, and multiple sentiment types introduce annotation noise.

1. Interpretability

The paper Contextual Sentiment Neural Network for Document Sentiment Analysis proposes a four‑layer interpretable model that computes word‑level raw sentiment scores, sentiment shift scores, contextual sentiment scores, and concept‑level sentiment scores, integrating dictionary features, negation handling, revised self‑attention, and k‑means clustering. The model uses an IP propagation method to make each layer correspond to interpretable sentiment components.

2. Contextual Aspect‑Sentiment Relationships

The study Weakly‑Supervised Aspect‑Based Sentiment Analysis via Joint Aspect‑Sentiment Topic Embedding introduces a topic‑model‑based weakly supervised method that learns joint aspect‑sentiment embeddings, pre‑trains a CNN with document‑topic similarity, and refines it via self‑training, achieving strong interpretability and clustering of aspect terms.

The paper Context‑aware Embedding for Targeted Aspect‑based Sentiment Analysis addresses the limitation of random target/aspect initialization by proposing a sparse coefficient vector to extract sentiment‑relevant words and fine‑tuning aspect vectors with contextual information, using the following objective functions:

X \times S

3. Noise in Labels

The work Learning with Noisy Labels for Sentence‑level Sentiment Classification proposes the NETAB model, consisting of two CNNs: one learns clean sentiment scores, the other learns a noise transition matrix. Training alternates between early pre‑training of the clean network and joint optimization, effectively handling up to 50% noisy labels.

4. Domain Sentiment Lexicon Construction

Several papers address lexicon building. Sentiment Lexicon Construction with Representation Learning Based on Hierarchical Sentiment Supervision learns sentiment‑aware word embeddings at both document and word levels, using seed words and a variant‑KNN classifier for expansion.

The method Automatic construction of domain‑specific sentiment lexicon based on constrained label propagation builds a word graph using WordNet, syntactic rules, and SOC‑PMI similarity, then propagates polarity from seed words to unlabeled words via semi‑supervised label propagation.

Conclusion

Across document‑level and aspect‑level sentiment analysis, the key research focus is discovering true sentiment contexts and modeling the semantic relationship between aspects and sentiment words, especially in domain‑specific data. Topic‑model‑based approaches are effective but rely on robust tokenization and entity recognition.

Future directions include transfer learning, mixup training, and semi‑supervised methods.

References

[1] Contextual Sentiment Neural Network for Document Sentiment Analysis [2] Weakly‑Supervised Aspect‑Based Sentiment Analysis via Joint Aspect‑Sentiment Topic Embedding [3] Context‑aware Embedding for Targeted Aspect‑based Sentiment Analysis [4] SentiLARE: Sentiment‑Aware Language Representation Learning with Linguistic Knowledge [5] Learning with Noisy Labels for Sentence‑level Sentiment Classification [6] Sentiment Lexicon Construction with Representation Learning Based on Hierarchical Sentiment Supervision [7] Automatic construction of domain‑specific sentiment lexicon based on constrained label propagation [8] STCS Lexicon: Spectral‑Clustering‑Based Topic‑Specific Chinese Sentiment Lexicon Construction for Social Networks [9] Sentiment Lexicon Construction with Hierarchical Supervision Topic Model