This thesis deals with solving multi-step tasks by using advice from experts, which are algorithms to solve individual steps of such tasks. We contribute with methods for maximizing the number of correct task solutions by selecting and combining experts for individual task instances and methods for automating the process of solving tasks on the Web, where experts are available as Web services.
Multi-step tasks frequently occur in Natural Language Processing (NLP) or Computer Vision, and as research progresses an increasing amount of exchangeable experts for the same steps are available on the Web. Service provider platforms such as Algorithmia monetize expert access by making expert services available via their platform and having customers pay for single executions.
Such experts can be used to solve diverse tasks, which often consist of multiple steps and thus require pipelines of experts to generate hypotheses.
We perceive two distinct problems for solving multi-step tasks with expert services: (1) Given that the task is sufficiently complex, no single pipeline generates correct solutions for all possible task instances. On ... mehre thus must learn how to construct individual expert pipelines for individual task instances in order to maximize the number of correct solutions, while also taking into account the costs adhered to executing an expert. (2) To automatically solve multi-step tasks with expert services, we need to discover, execute and compose expert pipelines. With mostly textual descriptions of complex functionalities and input parameters, Web automation entails to integrate available expert services and data, interpreting user-specified task goals or efficiently finding correct service configurations.
In this thesis, we present solutions to both problems: (1) We enable to learn well-performing expert pipelines assuming available reference data sets (comprising a number of task instances and solutions), where we distinguish between centralized and decentralized decision-making. We formalize the problem as specialization of a Markov Decision Process (MDP), which we refer to as Expert Process (EP) and integrate techniques from Statistical Relational Learning (SRL) or Multiagent coordination. (2) We develop a framework for automatically discovering, executing and composing expert pipelines by exploiting methods developed for the Semantic Web. We lift the representations of experts with structured vocabularies modeled with the Resource Description Framework (RDF) and extend EPs to Semantic Expert Processes (SEPs) to enable the data-driven execution of experts in Web-based architectures.
We evaluate our methods in different domains, namely Medical Assistance with tasks in Image Processing and Surgical Phase Recognition, and NLP for textual data on the Web, where we deal with the task of Named Entity Recognition and Disambiguation (NERD).