Computational Models for Determining the Personality Type of Text Authors

Osobine ličnosti opisuju dosljedne obrasce mišljenja, osjećanja i ponašanja ljudi. Postoji interes za automatsko određivanje osobina ličnosti autora iz teksta pošto se osobine ličnosti očituju u preferencama i izborima osobe tijekom života. Ovaj rad bavio se automatski određivanjem osobina ličnosti autora teksta na hrvatskom jeziku. Za potrebe rada prikupljen je skup podataka na hrvatskom jeziku koji se je sastoji od eseja te rezultata na upitniku ličnosti IPIP. Uz skup podataka na hrvatskom jeziku korišten je i skup podataka na engleskom jeziku te su vektori riječi na hrvatskom i engleskom jeziku poravnani u jedan vektorski prostor modelom MUSE. Potom su razvijena dva modela dubokog učenja za određivanje osobina ličnosti autora teksta. Prvi model je konvolucijska neuronska mreža, a drugi hijerarhijska mreža pažnje. Provedeno je i eksperimentalno vrednovanje modela i analiza modela modelom LIME.Personality traits describe consistent patterns of thoughts, feelings and behaviours. There is a continual interest in automated personality traits detection from texts, since personality traits are manifested in preferences and choices a person makes during their life. The focus of this paper has been in the automated personality detection from texts written in the Croatian language. For this purpose, a dataset was collected consisting of essays and IPIP personality questionary. Along with this dataset, an English language dataset was also used and the word embeddings for Croatian and English were aligned into one vector space using the MUSE model. In addition, two deep learning models for text-based personality traits detection were developed. The first model is a convolutional neural network and the second one is a hierarchical attention network. Experimental evaluation of the models and their analysis was done using the LIME model