This study proposes a framework for extracting unique discussions of the interests of managers and entrepreneurs on Twitter (X). By unique discussions of interests, we mean those that are more tweeted by these communities but rarely by public people. These discussions can be facts and/or sentiments related to some topics. Since this is a subjective problem, human intervention can lead to bias in the results. Therefore, we propose an unsupervised method with zero information about the context since prior knowledge stems from human intervention. Consequently, there is no real ground truth. To retrieve such discussions of interests, first, unique tweets (discussions) are identified in two stages. In the first stage, a scoring algorithm is proposed that gives a score to each tweet of a specific year and tweets are sorted based on their scores. Different sets of tweets are selected based on their scores and considered automatically created ground truths. In the next stage, an unsupervised convolutional neural network trained on the created ground truth is used for the classification of tweets of other years (whether they are unique to these communities). Finally, latent Dirichlet analysis is applied to the detected unique tweets to give the most common interest topics discussed by these communities. Experimental analysis is performed on tweets from 2017-2019. The results reveal these communities’ attitudes and highlight interesting common and different topics discussed between managers and entrepreneurs; some of them can be difficult for humans to predict in advance. The proposed approach is applicable to any community.