Semantic knowledge about concepts has been argued to be organized in two different ways: based on shared features (taxonomic) or based on co-occurrence in common scenes, events, or scenarios (thematic). Contemporary theories of semantic cognition, such as the dual-hub hypothesis (Mirman et al., 2017) and the controlled semantic cognition (CSC) framework (Ralph et al., 2017)assume that task-context influences which semantic systems are engaged, though competing theories differ in details about how this flexibility operates, for example, whether both taxonomic and thematic systems are equally engaged or disengaged (dual-hub) or whether the taxonomic organization is the core structure of semantic knowledge, with thematic organization flexibly engaged only in appropriate task settings (CSC). The first goal of the current study is to examine how flexibly the two semantic systems can be engaged under different task demands. To achieve this goal, my analyses largely rely on a multivariate analysis method called representational similarity analysis (RSA), a methodology that allows me to relate the pairwise similarity structure generated by neuroimaging data including EEG and fMRI to multiple computational models of taxonomic/thematic relations between concepts under different task contexts and make comparisons between models. The current project investigates datasets of a series of picture naming tasks with different semantic contexts. Across all task contexts, the similarity structure of neural activity correlated better with taxonomic than thematic measures in the time window of semantic processing. Most strikingly, patterns of neural activity between taxonomically related items were more similar to each than the patterns of neural activity for thematically related or unrelated items, even in tasks that focused attention on thematic relationships. At least in picture naming tasks, concepts are organized according to taxonomic knowledge in semantic space. These findings are not in line with the assumption task contexts influence the engagement of semantic systems indicated by the dual-hub theory or CSC account. The second part of my dissertation tests whether the taxonomic and thematic relations can be learned from texts using the same algorithm, focusing on different statistical regularities. Specifically, different types of word embeddings are learned through the same skip-gram model with negative sampling and their performances in measuring the strength of taxonomic and thematic relations were assessed based on human subjective ratings. The assessment results indicated that word2vec-based measures are essentially taxonomic measures. Although there is some flexibility in torquing the computational measure towards the thematic end by increasing the sliding window size within the same architecture of a skip-gram model, the effect is not large enough to make it an efficient thematic measure.