Providing coaching to participants in inquiry-based online discussions contributes to developing cognitive presence (CP) and higher-order thinking. However, a primary issue limiting quality and timely coaching is instructors' lack of tools to efficiently identify CP phases in massive discussion transcripts and effectively assess learners' cognitive development. This study examined a computational approach integrating text mining and co-occurrence analysis for assessing CP and cognitive development in online discussions based on the community of inquiry (CoI) framework. First, text classifiers trained on different language models were evaluated for identifying and coding the CP phases. Second, epistemic network analysis (ENA) was employed to model cognitive patterns reflected by co-occurrences between the coding elements. Results indicated that text classifiers trained on the state-of-the-art language model Bidirectional Encoder Representations from Transformers (BERT) can address the efficiency issue in coding CP phases in discussion transcripts and obtain substantial agreements (Cohen's "k" = 0.76) with humans, which outperformed other baseline classifiers. Furthermore, compared to traditional quantitative content analysis, ENA can effectively model the temporal characteristics of online discourse and detect fine-grained cognitive patterns. Overall, the findings suggest a feasible path for applying learning analytics to tracking learning progression and informing theory-based assessments.