Pipelines for Ad-hoc Large-scale Text Mining

Today's web search and big data analytics applications aim to address information needs~(typically given in the form of search queries) ad-hoc on large numbers of texts. In order to directly return relevant information instead of only returning potentially relevant texts, these applications have begun to employ text mining. The term text mining covers tasks that deal with the inference of structured high-quality information from collections and streams of unstructured input texts. Text mining requires task-specific text analysis processes that may consist of several interdependent steps. These processes are realized with sequences of algorithms from information extraction, text classification, and natural language processing. However, the use of such text analysis pipelines is still restricted to addressing a few predefined information needs. We argue that the reasons behind are three-fold: First, text analysis pipelines are usually made manually in respect of the given information need and input texts, because their design requires expert knowledge about the algorithms to be employed. When information needs have to be addressed that are unknown beforehand, text mining hence cannot be performed ad-hoc. Second, text analysis pipelines tend to be inefficient in terms of run-time, because their execution often includes analyzing texts with computationally expensive algorithms. When information needs have to be addressed ad-hoc, text mining hence cannot be performed in the large. And third, text analysis pipelines tend not to robustly achieve high effectiveness on all texts, because their results are often inferred by algorithms that rely on domain-dependent features of texts. Hence, text mining currently cannot guarantee to infer high-quality information. In this thesis, we contribute to the question of how to address information needs from text mining ad-hoc in an efficient and domain-robust manner. We observe that knowledge about a text analysis process and information obtained within the process help to improve the design, the execution, and the results of the pipeline that realizes the process. To this end, we apply different techniques from classical and statistical artificial intelligence. In particular, we first develop knowledge-based approaches for an ad-hoc pipeline construction and for an optimal execution of a pipeline on its input. Then, we show theoretically and practically how to optimize and adapt the schedule of the algorithms in a pipeline based on information in the analyzed input texts in order to maximize execution efficiency. Finally, we learn patterns in the argumentation structures of texts statistically that remain strongly invariant across domains and that, thereby, allow for more robust analysis results in a restricted set of tasks. We formally analyze all developed approaches and we implement them as open-source software applications. Based on these applications, we evaluate the approaches on established and on newly created collections of texts for scientifically and industrially important text analysis tasks, such as financial event extraction and fine-grained sentiment analysis. Our findings show that text analysis pipelines can be designed automatically, which process only portions of text that are relevant for the information need at hand. Through scheduling, the run-time efficiency of pipelines can be improved by up to more than one order of magnitude while maintaining effectiveness. Moreover, we provide evidence that a pipeline's domain robustness substantially benefits from focusing on argumentation structure in tasks like sentiment analysis. We conclude that our approaches denote essential building blocks of enabling ad-hoc large-scale text mining in web search and big data analytics applications.