Background: Traditionally text mention normalization corpora have normalized concepts to single ontology identifiers ("pre-coordinated concepts"). Less frequently, normalization corpora have used concepts with multiple identifiers ("post-coordinated concepts") but the additional identifiers have been restricted to a defined set of relationships to the core concept. This approach limits the ability of the normalization process to express semantic meaning. We generated a freely available corpus using post-coordinated concepts without a defined set of relationships that we term "compositional concepts" to evaluate their use in clinical text. Methods: We annotated 5397 disorder mentions from the ShARe corpus to SNOMED CT that were previously normalized as "CUI-less" in the "SemEval-2015 Task 14" shared task because they lacked a pre-coordinated mapping. Unlike the previous normalization method, we do not restrict concept mappings to a particular set of the Unified Medical Language System (UMLS) semantic types and allow normalization to occur to multiple UMLS Concept Unique Identifiers (CUIs). We computed annotator agreement and assessed semantic coverage with this method. Results: We generated the largest clinical text normalization corpus to date with mappings to multiple identifiers and made it freely available. All but 8 of the 5397 disorder mentions were normalized using this methodology. Annotator agreement ranged from 52.4% using the strictest metric (exact matching) to 78.2% using a hierarchical agreement that measures the overlap of shared ancestral nodes. Conclusion: Our results provide evidence that compositional concepts can increase semantic coverage in clinical text. To our knowledge we provide the first freely available corpus of compositional concept annotation in clinical text.

In this research the focus is on data-centric AI with a specific concentration on data quality evaluation and improvement for machine learning. We first present a practical framework for data quality evaluation and improvement, using a legal domain as a case study and build a corpus for legal argument mining. We first created an initial corpus with 4,937 instances that were manually labeled. We define five data quality evaluation dimensions: comprehensiveness, correctness, variety, class imbalance, and duplication, and conducted a quantitative evaluation on these dimensions for the legal dataset and two existing datasets in the medical domain for medical concept normalization. The first group of experiments showed that class imbalance and insufficient training data are the two major data quality issues that negatively impacted the quality of the system that was built on the legal corpus. The second group of experiments showed that the overlap between the test datasets and the training datasets, which we defined as "duplication," is the major data quality issue for the two medical corpora. We explore several widely used machine learning methods for data quality improvement. Compared to pseudo-labeling, co-training, and expectation-maximization (EM), generative adversarial network (GAN) is more effective for automated data augmentation, especially when a small portion of labeled data and a large amount of unlabeled data is available. The data validation process, the performance improvement strategy, and the machine learning framework for data evaluation and improvement discussed in this dissertation can be used by machine learning researchers and practitioners to build high-performance machine learning systems. All the materials including the data, code, and results will be released at: https://github.com/haihua0913/dissertation-dqei.