The main goal of this thesis is to understand the topological properties of semantic networks, to find language-specific patterns, and to investigate their connection principles. Interpreting unstructured texts in natural language is a crucial task for computers. Natural Language Processing (NLP) applications rely on semantic networks for structured knowledge representation. Although NLP technologies have been applied to various domains with some degree of success, they still face many challenges due to the ambiguity of human language. To inform better algorithms, we need to pay attention to fundamental structures of semantic networks in different languages. However, these remain to be investigated properly. In this thesis we extract semantic networks with 7 distinct relations for 11 languages from ConceptNet. We systematically analyze the degree distribution, degree correlation and clustering of these networks. We also measure their structural similarity and complementarity coefficients. Our findings show that semantic networks have universalities in basic structures: they have high sparsity, high clustering, and power-law degree distributions. Our findings also show that the majority of the considered networks are scale-free. In addition, our results show that networks in different languages exhibit different properties, which are determined by grammatical rules. For example, the networks of highly inflected languages show peaks in the degree distributions that deviate from a power-law. Furthermore, we find that depending on the type of semantic relation and the language, the connection principles of networks are different. Some networks are more similarity-based, while others are more complementarity-based. We conclude the thesis by demonstrating how the knowledge of similarity and complementarity can better inform NLP in link prediction tasks.

Online searching for healthcare information has gradually become a widely used internet case. Suppose a patient suffers the symptom but is unsure of the action he needs to take, a self-diagnosis tool can help the patient identify the possible conditions and whether this patient needs to seek immediate medical help. However, the accuracy and quality of the service provided by those self-diagnosis tools are still disappointing and need further improvement. This thesis focuses on an automatic differential diagnosis task with a comprehensive evaluation of reinforcement learning methods. Also, we present a systematic method to simulate medically correct patients records, which integrates a standard symptom modeling approach called NLICE. In this way, we can bridge the gap between limited available patients records and data-driven healthcare methodologies. This project investigates both flat-RL methods and hierarchical RL in an automatic differential diagnosis setting and evaluates the performance of those two kinds of methods on simulated patients records. More specifically, the action space for the differential diagnosis task is inevitably large, so the flat-RL performs relatively poorly in complicated scenarios. The hierarchical RL method can split a complex diagnosis task into smaller tasks: it contains two-level of policy learning, and each low-layer policy imitates one medical specialty. Therefore hierarchical RL method increases the Top 1 success rate from 23.1\% in flat-RL method to 45.4\%.

Besides the advanced policy learning strategy, this thesis explores the ability of NLICE symptom modeling in distinguishing conditions that share similar symptoms. The experimental results experience increases in flat-RL and hierarchical RL models and finally achieve 36.2\% and 71.8\% Top 1 success rates, respectively. To further solve the sparse action space problem in the automatic diagnosis domain, the reward shaping algorithm is implemented in the reward configuration part. The average gained reward of hierarchical RL increases from -3.65 to 0.87. Additionally, we model the general demographic background of patients and utilize contextual information to perform the policy transformation strategy, which eliminates the miss classification problem in highly sex-age related diseases.

Networks are all around us, telecommunication networks, road transportation networks, and the Internet are a few examples of networks that we encounter every day. The entities in a network are represented by nodes and the interconnections between them are represented by links. For example, in a telecommunication network, a node could be an end point for data transmission, a redistribution point or in physical terms, an entity that is capable of receiving, transmitting or redistributing information and a link could be a wired or a wireless connection between the nodes. It is of prime importance that the networks perform their functions properly. To ensure the effective operation of such networks, we need to control them by applying external inputs on some nodes which are known as driver nodes. We say that a network is controllable if it can be driven from any arbitrary state to any desired state in finite time under the control of driver nodes which are attached to the external inputs. Networks are often confronted with perturbations in the form of targeted and random attacks to disrupt their operation. Such perturbations make these networks less controllable. Thus, more driver nodes are needed to gain the full controllability of these networks. As a result, the robustness of network controllability decreases. In this study, the minimum number of driver nodes which gain full controllability after failures or attacks is chosen as the robustness metric.
Existing closed-form analytical approximations estimate the normalized minimum number of driver nodes as a function of the fraction of removed links in both targeted and random attacks. However, the approximations sometimes do not fit with the simulations and the errors between the approximations and simulations are large. The main objectives of this study are to improve the analytical approximations using machine learning methods for both targeted and random attacks and additionally, derive a suitable analytical approximation for the out-in degree-based attack. Specifically, we use Linear Regression, Random Forest, and Artificial Neural Networks. To evaluate the performance of our machine learning models, we compare them with analytical approximations and simulations. In addition to this, we also study the attack based variability in estimating the minimum number of driver nodes using robustness envelopes.
Based on the performance evaluations, we found that the approximations using machine learning models significantly outperform the existing closed-form analytical approximations for the minimum number of driver nodes, both in real-world and synthetic networks. Furthermore, we also assess the performance of our analytical approximations for the out-in degree-based attacks by comparing them with simulations.