Abstract

The strength and power of structured prediction approaches in machine learning originates from a proper recognition and exploitation of inherent structural dependencies within complex objects, which structural models are trained to output. Among the complex tasks that benefited from structured prediction approaches, clustering is of a special interest. Structured output models based on representing clusters by latent graph structures made the task of supervised clustering tractable. While in practice these models proved effective in solving the complex NLP task of coreference resolution, in this thesis, we aim at exploring their capacity to be extended to other tasks and domains, as well as the methods for performing such adaptation and for improvement in general, which, as a result, go beyond clustering and are commonly applicable in structured prediction. Studying the extensibility of the structural approaches for supervised clustering, we apply them to two different domains in two different ways. First, in the networking domain, we do clustering of network traffic by adapting the model, taking into account the continuity of incoming data. Our experiments demonstrate that the structural clustering approach is not only effective in such a scenario, but also, if changing the perspective, provides a novel potentially useful tool for detecting anomalies. The other part of our work is dedicated to assessing the amenability of the structural clustering model to joint learning with another structural model, for ranking. Our preliminary analysis in the context of the task of answer-passage reranking in question answering reveals a potential benefit of incorporating auxiliary clustering structures. Due to the intrinsic complexity of the clustering task and, respectively, its evaluation scenarios, it gave us grounds for studying the possibility and the effect from optimizing task-specific complex measures in structured prediction algorithms. It is common for structured prediction approaches to optimize surrogate loss functions, rather than the actual task-specific ones, in or- der to facilitate inference and preserve efficiency. In this thesis, we, first, study when surrogate losses are sufficient and, second, make a step towards enabling direct optimization of complex structural loss functions. We propose to learn an approximation of a complex loss by a regressor from data. We formulate a general structural framework for learning with a learned loss, which, applied to a particular case of a clustering problem – coreference resolution, i) enables the optimization of a coreference metric, by itself, having high computational complexity, and ii) delivers an improvement over the standard structural models optimizing simple surrogate objectives. We foresee this idea being helpful in many structured prediction applications, also as a means of adaptation to specific evaluation scenarios, and especially when a good loss approximation is found by a regressor from an induced feature space allowing good factorization over the underlying structure.