Deep learning methods for knowledge base population

Adel, Heike (2018): Deep learning methods for knowledge base population. Dissertation, LMU München: Fakultät für Mathematik, Informatik und Statistik

Abstract

Knowledge bases store structured information about entities or concepts of the world and can be used in various applications, such as information retrieval or question answering. A major drawback of existing knowledge bases is their incompleteness. In this thesis, we explore deep learning methods for automatically populating them from text, addressing the following tasks: slot filling, uncertainty detection and type-aware relation extraction. Slot filling aims at extracting information about entities from a large text corpus. The Text Analysis Conference yearly provides new evaluation data in the context of an international shared task. We develop a modular system to address this challenge. It was one of the top-ranked systems in the shared task evaluations in 2015. For its slot filler classification module, we propose contextCNN, a convolutional neural network based on context splitting. It improves the performance of the slot filling system by 5.0% micro and 2.9% macro F1. To train our binary and multiclass classification models, we create a dataset using distant supervision and reduce the number of noisy labels with a self-training strategy. For model optimization and evaluation, we automatically extract a labeled benchmark for slot filler classification from the manual shared task assessments from 2012-2014. We show that results on this benchmark are correlated with slot filling pipeline results with a Pearson's correlation coefficient of 0.89 (0.82) on data from 2013 (2014). The combination of patterns, support vector machines and contextCNN achieves the best results on the benchmark with a micro (macro) F1 of 51% (53%) on test. Finally, we analyze the results of the slot filling pipeline and the impact of its components. For knowledge base population, it is essential to assess the factuality of the statements extracted from text. From the sentence "Obama was rumored to be born in Kenya", a system should not conclude that Kenya is the place of birth of Obama. Therefore, we address uncertainty detection in the second part of this thesis. We investigate attention-based models and make a first attempt to systematize the attention design space. Moreover, we propose novel attention variants: External attention, which incorporates an external knowledge source, k-max average attention, which only considers the vectors with the k maximum attention weights, and sequence-preserving attention, which allows to maintain order information. Our convolutional neural network with external k-max average attention sets the new state of the art on a Wikipedia benchmark dataset with an F1 score of 68%. To the best of our knowledge, we are the first to integrate an uncertainty detection component into a slot filling pipeline. It improves precision by 1.4% and micro F1 by 0.4%. In the last part of the thesis, we investigate type-aware relation extraction with neural networks. We compare different models for joint entity and relation classification: pipeline models, jointly trained models and globally normalized models based on structured prediction. First, we show that using entity class prediction scores instead of binary decisions helps relation classification. Second, joint training clearly outperforms pipeline models on a large-scale distantly supervised dataset with fine-grained entity classes. It improves the area under the precision-recall curve from 0.53 to 0.66. Third, we propose a model with a structured prediction output layer, which globally normalizes the score of a triple consisting of the classes of two entities and the relation between them. It improves relation extraction results by 4.4% F1 on a manually labeled benchmark dataset. Our analysis shows that the model learns correct correlations between entity and relation classes. Finally, we are the first to use neural networks for joint entity and relation classification in a slot filling pipeline. The jointly trained model achieves the best micro F1 score with a score of 22% while the neural structured prediction model performs best in terms of macro F1 with a score of 25%.

• Dokumententyp: Dissertationen (Dissertation, LMU München)
• Keywords: knowledge base population, information extraction, deep learning
• Themengebiete: 000 Allgemeines, Informatik, Informationswissenschaft ; 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik
• Fakultäten: Fakultät für Mathematik, Informatik und Statistik
• Sprache der Hochschulschrift: Englisch
• Datum der mündlichen Prüfung: 26. Juni 2018
• 1. Bericht­erstatter:in: Schütze, Hinrich