Learning Efficient Information Extraction on Heterogeneous Texts

International Joint Conference on Natural Language Processing, pages 534–542, Nagoya, Japan, 14-18 October 2013. Learning Efficient Information Extraction on Heterogeneous Texts Henning Wachsmuth University of Paderborn, s-lab Paderborn, Germany hwachsmuth@s-lab.upb.de Benno Stein Bauhaus-Universit¨at Weimar Weimar, Germany benno.stein@uni-weimar.de Gregor Engels University of Paderborn, s-lab Paderborn, Germany engels@upb.de Abstract From an efficiency viewpoint, information extraction means to filter the relevant por- tions of natural language texts as fast as possible. Given an extraction task, differ- ent pipelines of algorithms can be devised that provide the same precision and recall but that vary in their run-time due to dif- ferent pipeline schedules. While recent re- search investigated how to determine the run-time optimal schedule for a collection or a stream of texts, this paper goes one step beyond: we analyze the run-times of efficient schedules as a function of the het- erogeneity of the texts and we show how this heterogeneity is characterized from a data perspective. For extraction tasks on heterogeneous big data, we present a self- supervised online adaptation approach that learns to predict the optimal schedule de- pending on the input text. Our evaluation suggests that the approach will significant- ly improve efficiency on collections and streams of texts of high heterogeneity. 1 Introduction Information extraction analyzes natural language text in order to find relevant information about en- tities and the events they participate in. An extrac- tion task often requires to fill event templates with considerable numbers of slots. Such a task implies several analysis steps, e.g. certain types of entity and relation extraction, and it is therefore typically tackled with a pipeline Π = ⟨A, π⟩, where A is a set of extraction algorithms and π a schedule that prescribes the order of algorithm application. The information sought for is anchored in text units of a certain size, e.g. in a sentence or paragraph. In times of big data, the run-time efficiency of information extraction receives much attention in research and industry (Chiticariu et al., 2010). Among others, a growing need for business intelli- gence can be regarded as the driving force behind. This trend is equally observed by consulting com- panies who see evolving markets for predictive an- alytics (Harper, 2011), by global software players who exploit big data for decision making (White, 2011), and by researchers who seek to annotate ta- bles at web scale (Limaye et al., 2010). Generally, a pipeline Π = ⟨A, π⟩ can be sped up by parallelization (Agichtein, 2005), if given enough resources, or by using faster but less ef- fective algorithms (Al-Rfou’ and Skiena, 2012). In addition, information extraction can always be approached as a filtering task as discussed in detail in (Wachsmuth et al., 2013b): By filling a template slot, each algorithm in A implicitly classifies cer- tain units of an input text as relevant. Only these units need to be filtered for the next algorithm in π. As a result, a smart schedule π will often signif- icantly improve the overall extraction efficiency. If the input requirements of …

Content truncated to 500 words