چكيده به لاتين
Nowadays, with the expansion of social networks, the way people communicate with each other has
changed, offering a new perspective on human behavior. The personality of each user can be indicative of
their behavior and interests. Personality is a combination of behavior, emotions, motivations, individual
traits, and thought patterns that influence one's lifestyle, well-being, health, preferences, and desires.
Therefore, the ability to recognize personality traits has many important practical applications. Some of
these applications include advanced personal assistants, recommendation systems, advertisements, and
using personality traits in job interviews. Psychologists typically use questionnaires directly to determine
individuals' personality types, but individuals' information such as their texts and images on social
networks can also indicate their personality types.
In this research, three types of data, textual, visual, and audio, are used to identify and predict individuals'
personality traits. Therefore, the proposed model in this project is a three-mode deep learning-based
model, which generally consists of four components. In the textual component, the proposed model
uses transformer-based language models. In the visual component, a spatio-temporal model based on
pre-trained models on a significant number of images and transformer encoder layers is proposed, and
the impact of individuals' images on their personality traits is measured. In the audio component,
audio features are extracted using deep networks. In the final four-mode model, pre-trained singlemode models are first used to extract single-mode features. In this research, transformer layers are
suggested for combining single-mode features, which significantly improve the performance of the multimode model. Additionally, a unified-bert model is proposed for the first time in this problem, which is
used to extract visual-linguistic features.
The proposed models in this research achieved promising results in identifying individuals' personalities,
which showed a significant improvement compared to recent work in this field. The four-mode model
proposed in this research achieved an accuracy of 92%, which improved the accuracy by about 0.8%
compared to the 2022 research. Furthermore, in the correlation coefficient of alignment metric introduced
in recent years, our model improved it by 12% compared to the 2022 research, which is a significant
improvement. Various combinations of models were also tested in this research, and the results of all
these models also showed significant improvements compared to recent research.
