Textual contents are very powerful sources of information about events. Twitter, as a microblogging social network provides such a platform for users in order to publish their thoughts about a special happening. However, deliberation on this social network indicates that words are majorly used in sentiments different that their original sounds. For instance, the word Damn or Hell despite their negative sounds are widely used to express positive feelings about the subject. In addition, there is a wide variety of words which are used in so-called microblogging form, rather than their original format which hinders the use of current sentimental databases to return the meaning. Lack of such database, makes the process of acquiring information with difficulties. The current paper introduces a novel topic modeling method for finding the sentiments in short texts using previous knowledge of trusted sources for sentiments. As the trusted source of sentiments, we utilized emojis that were extracted from a database of nearly 8 million tweets. The included in the process of sentiment analysis in a process of 9 stages. Further, the acquired word in each stage were appended to the next stage in form of a back propagation algorithm in order to incorporate the proposed method over the large sets of short texts which can be applied in real tweets. Thus, a sentimental database for heterogeneous words distinguishing sentimentally positive and negative words, and including hidden words named as Golden List were created. A validation test over the words acquired in each step, shows a sign of higher occurrence rate and sentimental closeness in negative top words rather than in positive top words in terms of Jaccard similarity. Moreover, the comparisons of the created dataset with two databases with the maximum of 20% word hit, indicates similarity rate of above 55% for both positive and negative top words.