Classification of staff demands in the public institutions by artificial intelligence methods

The global COVID-19 pandemic has rapidly increased the use of information technologies worldwide. Nowadays, public institutions and private businesses receive requests from users (customers, employees, etc.) through various "Job Demand System" software applications over the Internet. The extensive use of information technologies results in numerous demands for departments of public institutions or private businesses, which means processing and transforming large-scale data into information. Artificial intelligence methods provide significant advantages in making accurate inferences from big data. This thesis aims to automate the determination of which department the job request correspondence belongs to using artificial intelligence technology. Thus, it aims to reduce errors that may occur in the manual determination process of the relevant unit in job request systems and to prevent time and energy losses. A dataset was initially obtained since no dataset related to job demand software was found in the literature. The dataset used in this thesis includes 20.602 job requests sent by personnel in a public institution. Then, four different deep learning-based hybrid methods were created using the LSTM architecture and the word embedding methods BERTurk, Word2Vec, Doc2Vec, and FastText to estimate the job requests belonging to which department. Pre-processing steps such as cleaning and/or stemming were applied to the datasets to improve the success levels of the hybrid methods. Finally, the hybrid methods were compared to each other. In the experimental results, LSTM and Word2Vec, LSTM and Doc2Vec, and LSTM and FastText methods classified job demand with an average success rate of 92%. The LSTM and BERTurk method showed lower success (88%) than others. Only the cleaning step among the pre-processing steps improved the success of the LSTM and BERTurk hybrid method by 2%. It was observed that applying cleaning and stemming or only stemming did not improve the success of this hybrid method. In other methods, applying both pre-processing steps together increased the success rate by 1%. As a result, in this thesis, the job demand system software was classified accurately according to the units they may be relevant to in a public institution.