Virtual assistant design for analysis of Turkish social media contents

With the rapid development of technology in recent years, people have started to use social media platforms as a means of socialization. In social media platforms, people are both in the position of producing and consuming information. This information may be about the product, service, or agenda received. In addition to individual users, commercial platforms can use social media to reach their target audiences and manage their brand perceptions. Social media analytics comes into play at this point and enables the creation of commercial and individual benefits by analyzing the patterns in the unstructured data in social media. In the background of social media analytics, there are text mining processes. Since social media data is speech-based, it contains structures such as abbreviations and emojis and requires a more sensitive process than text mining. In this study, a program interface was studied on the Twitter platform, which is frequently used in social media analysis, that enables data collection, preprocessing and analysis processes to be carried out with instructions. Through this interface, the mentioned social media analytics process can be carried out step by step over Twitter data. Here, researchers can start the analysis process by shooting Tweets for any topic, product or service they want, in the date range, language and number they want. The researcher is guided by being informed about the size of the data desired to be obtained before the data extraction process. In the data preprocessing step, data unitization, standardization and filtering can be performed with various orientations. For example, in the standardization phase, the numerical distribution of the special characters in the data is given to the researcher. In this way, the researcher can clean the desired characters or transform phrases in the spoken language such as emoji and abbreviations. In the analysis phase, there are sentiment analysis and topical modeling options. The Turkish-BERT model was used for the sentiment analysis process. In topic modeling, machine learning processes are performed on the labeled category data, and the most successful model is selected and classification is carried out with this model. It is aimed to make the program interface supported by machine learning in future studies by recording the operations performed during the entire analysis process, analysis results and data detail information.