Words in a document are sometimes classified as objective (delivering facts) or subjective (expressing ideas) based on the context in which they appear. If we have a collection of papers on the order Hemiptera, and someone assigns the term "bug" to that subject, it likely refers to a specific kind of insect, but when used to the topic of "software," it likely communicates a negative view. In this research, we propose a model called discriminatively objective-subjective LDA (dosLDA) based on the intuitive premise that various words have varied degrees of discriminative strength in providing the objective meaning or the subjective sense with regard to their assigned themes. In order to capture the relationship between themes and discriminative power for the words in a document in a supervised way, the proposed dosLDA makes use of a pair of objective and subjective selection factors. Therefore, each text should be shown as a "bag-of-discriminatory-words" (BoDW). Experiments with both textual and visual data show that dosLDA not only outperforms conventional methods in terms of topic modeling and document categorization, but can also determine if a given word has a more objective or subjective meaning in relation to its assigned topic.