Neural information extraction from natural language text

Gupta, Pankaj (2019): Neural information extraction from natural language text. Dissertation, LMU München: Fakultät für Mathematik, Informatik und Statistik

Abstract

Natural language processing (NLP) deals with building computational techniques that allow computers to automatically analyze and meaningfully represent human language. With an exponential growth of data in this digital era, the advent of NLP-based systems has enabled us to easily access relevant information via a wide range of applications, such as web search engines, voice assistants, etc. To achieve it, a long-standing research for decades has been focusing on techniques at the intersection of NLP and machine learning. In recent years, deep learning techniques have exploited the expressive power of Artificial Neural Networks (ANNs) and achieved state-of-the-art performance in a wide range of NLP tasks. Being one of the vital properties, Deep Neural Networks (DNNs) can automatically extract complex features from the input data and thus, provide an alternative to the manual process of handcrafted feature engineering. Besides ANNs, Probabilistic Graphical Models (PGMs), a coupling of graph theory and probabilistic methods have the ability to describe causal structure between random variables of the system and capture a principled notion of uncertainty. Given the characteristics of DNNs and PGMs, they are advantageously combined to build powerful neural models in order to understand the underlying complexity of data. Traditional machine learning based NLP systems employed shallow computational methods (e.g., SVM or logistic regression) and relied on handcrafting features which is time-consuming, complex and often incomplete. However, deep learning and neural network based methods have recently shown superior results on various NLP tasks, such as machine translation, text classification, namedentity recognition, relation extraction, textual similarity, etc. These neural models can automatically extract an effective feature representation from training data. This dissertation focuses on two NLP tasks: relation extraction and topic modeling. The former aims at identifying semantic relationships between entities or nominals within a sentence or document. Successfully extracting the semantic relationships greatly contributes in building structured knowledge bases, useful in downstream NLP application areas of web search, question-answering, recommendation engines, etc. On other hand, the task of topic modeling aims at understanding the thematic structures underlying in a collection of documents. Topic modeling is a popular text-mining tool to automatically analyze a large collection of documents and understand topical semantics without actually reading them. In doing so, it generates word clusters (i.e., topics) and document representations useful in document understanding and information retrieval, respectively. Essentially, the tasks of relation extraction and topic modeling are built upon the quality of representations learned from text. In this dissertation, we have developed task-specific neural models for learning representations, coupled with relation extraction and topic modeling tasks in the realms of supervised and unsupervised machine learning paradigms, respectively. More specifically, we make the following contributions in developing neural models for NLP tasks: 1. Neural Relation Extraction: Firstly, we have proposed a novel recurrent neural network based architecture for table-filling in order to jointly perform entity and relation extraction within sentences. Then, we have further extended our scope of extracting relationships between entities across sentence boundaries, and presented a novel dependency-based neural network architecture. The two contributions lie in the supervised paradigm of machine learning. Moreover, we have contributed in building a robust relation extractor constrained by the lack of labeled data, where we have proposed a novel weakly-supervised bootstrapping technique. Given the contributions, we have further explored interpretability of the recurrent neural networks to explain their predictions for the relation extraction task. 2. Neural Topic Modeling: Besides the supervised neural architectures, we have also developed unsupervised neural models to learn meaningful document representations within topic modeling frameworks. Firstly, we have proposed a novel dynamic topic model that captures topics over time. Next, we have contributed in building static topic models without considering temporal dependencies, where we have presented neural topic modeling architectures that also exploit external knowledge, i.e., word embeddings to address data sparsity. Moreover, we have developed neural topic models that incorporate knowledge transfers using both the word embeddings and latent topics from many sources. Finally, we have shown improving neural topic modeling by introducing language structures (e.g., word ordering, local syntactic and semantic information, etc.) that deals with bag-of-words issues in traditional topic models. The class of proposed neural NLP models in this section are based on techniques at the intersection of PGMs, deep learning and ANNs. Here, the task of neural relation extraction employs neural networks to learn representations typically at the sentence level, without access to the broader document context. However, topic models have access to statistical information across documents. Therefore, we advantageously combine the two complementary learning paradigms in a neural composite model, consisting of a neural topic and a neural language model that enables us to jointly learn thematic structures in a document collection via the topic model, and word relations within a sentence via the language model. Overall, our research contributions in this dissertation extend NLP-based systems for relation extraction and topic modeling tasks with state-of-the-art performances.

• Dokumententyp: Dissertationen (Dissertation, LMU München)
• Keywords: Information Extraction, Natural Language Processing, Relation Extraction, Named Entity Extraction, Topic Modeling, Document Representation, Neural Networks, Transfer Learning, Unsupervised deep learning, Generative modeling, Neural Composite modeling, Textual Similarity, Interpretability of Neural Networks
• 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. September 2019
• 1. Bericht­erstatter:in: Schütze, Hinrich