The project has collected over 300K labelled and unlablled patient reviews for model development. The keywords in the collection have been mapped to the aspects of health care services. On the constructed dataset, the project has proposed and implemented several novel approaches with 12 peer-reviewed publication in top-tier conferences or high impact factor journals including:
1) A novel framework based on reinforcement learning for topic modelling in order to improve the topic coherence measures. The experimental result on the patient reviews shows that it beats several existing topic models in generating more coherent topics. The technical training for the Fellow focused on the knowledge of reinforcement learning. The work has been published in EMNLP 2019, which is one of the top conferences in Natural Language Processing (NLP).
2) A convolutional attention-based neural sentiment model for clinical text classification. The experimental result shows that the proposed method beats several strong baselines on the collected dataset. The visualization of learned representation also gives a better interpretation of embedding space. The training in this WP for the Fellow includes data visualisation. The work has been published in IEEE ACCESS (impact factor: 3.557).
3) A joint learning method with Generative Adversarial Network (GAN) which improves the performance both in aspect extraction and sentiment classification compared with the neural topic model and neural sentiment model. It improved the performance both in aspect extraction and sentiment classification compared with the neural topic model and neural sentiment model. The work has been accepted by IEEE Transactions on Knowledge and Data Engineering (impact factor: 4.935).
4) A weakly-supervised model which can automatically discover new words describing patient experience through the statistical learning process and simultaneously capture aspect-level opinions on clinical text. It improved the performance both in aspect extraction and sentiment classification compared with the pre-trained language neural topic model and neural sentiment model. This work has been submitted to Artificial Intelligence In Medicine. (Impact factor: 4.383)
5) An Adversarial Multi-task Learning Framework to identify the aspect-invariant/dependent sentiment expressions automatically without requiring extra annotations. Experimental results on two benchmark datasets show that extending existing neural models using our proposed framework achieves superior performance. In addition, the aspect-invariant data extracted by our framework can be considered as pivot features for better transfer learning of the ABSA models on unseen aspects. This work has been accepted by CIKM 2020, which is one of the top-tier conferences in Data Mining and Information Retrieval.
All of the proposed methods above are evaluated on public dataset or project collected dataset and achieve a state of the art performance. Besides, we also implement the proposed method on relevant natural language processing and machine learning application such as emotion cause extraction, question answering, binding prediction on DNA sequence, and stance detection on social media. Most of those works appear in top-tier conferences or high-impact factor journals such as ACL, EMNLP, COLING, and Bioinformatics.
