This thesis introduces a robust end to end topic discovery framework that extracts a set of coherent topics stemming intrinsically from document similarities. Some topic clustering methods can support embedded vectors instead of traditional Bag-of-Words (BoW) representation. Some can be free from the number of topics hyperparameter and some others can extract a multi-scale relation between topics. However, no topic clustering method supports all these properties together. This thesis focuses on this gap in the literature by designing a framework that supports any type of document-level features especially the embedded vectors. This framework does not require any uninformed decision making about the underlying data such as the number of topics, instead, the framework extracts topics in multiple resolutions. To achieve this goal, we combine existing methods from natural language processing (NLP) for feature generation and graph theory, first for graph construction based on semantic document similarities, then for graph partitioning to extract corresponding topics in multiple resolutions. Finally, we use specific methods from statistical machine learning to obtain highly generalisable supervised models to deploy topic classifiers for the deployment of topic extraction in real-time. Our applications on both a noisy and specialised corpus of medical records (i.e., descriptions for patient incidents within the NHS) and public news articles in daily language show that our framework extracts coherent topics that have better quantitative benchmark scores than other methods in most cases. The resulting multi-scale topics in both applications enable us to capture specific details more easily and choose the relevant resolutions for the specific objective. This study contributes to topic clustering literature by introducing a novel graph theoretical perspective that provides a combination of new properties. These properties are multiple resolutions, independence from uninformed decisions about the corpus, and usage of recent NLP features, such as vector embeddings.