| Přispěvatelé: |
Schommer, Christoph [superviser], van der Torre, Leon [president of the jury], Ziafati, Pouyan [member of the jury], Dong, Tiansi [member of the jury], Hui, Kai [member of the jury] |
| Popis: |
People have different lexical choices when expressing their opinions. Sentiment analysis, as a way to automatically detect and categorize people’s opinions in text, needs to reflect this diversity. In this research, I look beyond the traditional population-level sentiment modeling and leverage socio-psychological theories to incorporate the concept of personalized modeling. In particular, a hierarchical neural network is constructed, which takes related information from a person’s past expressions to provide a better understanding of the sentiment from the expresser’s perspective. Such personalized models can suffer from the data sparsity issue, therefore they are difficult to develop. In this work, this issue is addressed by introducing the user information at the input such that the individuality from each user can be captured without building a model for each user and the network is trained in one process. The evolution of a person’s sentiment over time is another aspect to investigate in personalization. It can be suggested that recent incidents or opinions may have more effect on the person’s current sentiment than the older ones, and the relativeness between the targets of the incidents or opinions plays a role on the effect. Moreover, psychological studies have argued that individual variation exists in how frequently people change their sentiments. In order to study these phenomena in sentiment analysis, an attention mechanism which is reshaped with the Hawkes process is applied on top of a recurrent network for a user-specific design. Furthermore, the modified attention mechanism delivers a functionality in addition to the conventional neural networks, which offers flexibility in modeling information decay for temporal sequences with various time intervals. The developed model targets data from social platforms and Twitter is used as an example. After experimenting with manually and automatically labeled datasets, it can be found that the input formulation for representing the concerned information and the network design are the two major impact factors of the performance. With the proposed model, positive results have been observed which confirm the effectiveness of including user-specific information. The results reciprocally support the psychological theories through the real-world actions observed. The research carried out in this dissertation demonstrates a comprehensive study of the significance of considering individuality in sentiment analysis, which opens up new perspectives for future research in the area and brings opportunities for various applications. |
