Relation extraction (RE) is a task within natural language processing focused on the classification of relationships between entities in a given text. Primary applications of RE can be seen in various contexts such as knowledge graph construction and question answering systems. Traditional approaches to RE tend towards the prediction of relationships between exactly two entity mentions in small text snippets. However, with the introduction of datasets such as DocRED, research in this niche has progressed into examining RE at the document-level. Document-level relation extraction (DocRE) disrupts conventional approaches as it inherently introduces the possibility of multiple mentions of each unique entity throughout the document along with a significantly higher probability of multiple relationships between entity pairs. There have been many effective approaches to document-level RE in recent years utilizing various architectures, such as transformers and graph neural networks. However, all of these approaches focus on the classification of a fixed number of known relationships. As a result of the large quantity of possible unique relationships in a given corpus, it is unlikely that all interesting and valuable relationship types are labeled before hand. Furthermore, traditional naive approaches to clustering on unlabeled data to discover novel classes are not effective as a result of the unique problem of large true negative presence. Therefore, in this work we propose a multi-step filter and train approach leveraging the notion of contrastive representation learning to discover novel relationships at the document level. Additionally, we propose the use of an alternative pretrained encoder in an existing DocRE solution architecture to improve F1 performance in base multi-label classification on the DocRED dataset by 0.46. To the best of our knowledge, this is the first exploration of novel class discovery applied to the document-level RE task. Based upon our holdout evaluation method, we increase novel class instance representation in the clustering solution by 5.5 times compared to the naive approach and increase the purity of novel class clusters by nearly 4 times. We then further enable the retrieval of both novel and known classes at test time provided human labeling of cluster propositions achieving a macro F1 score of 0.292 for novel classes. Finally, we note only a slight macro F1 decrease on previously known classes from 0.402 with fully supervised training to 0.391 with our novel class discovery training approach.

Nowadays, the effective management of information is extremely important for all spheres of our lives and applications such as search engines and question answering systems help users to find the information that they need. However, even when assisted by these various applications, people sometimes struggle to find what they want. For example, when choosing a product customers can be confused by the need to consider many features before they can reach a decision. Interactive question answering (IQA) systems can help customers in this process, by answering questions about products and initiating a dialogue with the customers when their needs are not clearly defined. The focus of this thesis is how to design an interactive question answering system that will assist users in choosing a product they are looking for, in an optimal way, when a large number of similar products are available. Such an IQA system will be based on selecting a set of characteristics (also referred to as product features in this thesis), that describe the relevant product, and narrowing the search space. We believe that the order in which these characteristics are presented in terms of these IQA sessions is of high importance. Therefore, they need to be ranked in order to have a dialogue which selects the product in an efficient manner. The research question investigated in this thesis is whether product characteristics mentioned in user reviews are important for a person who is likely to purchase a product and can therefore be used when designing an IQA system. We focus our attention on products such as mobile phones; however, the proposed techniques can be adapted for other types of products if the data is available. Methods from natural language processing (NLP) fields such as coreference resolution, relation extraction and opinion mining are combined to produce various rankings of phone features. The research presented in this thesis employs two corpora which contain texts related to mobile phones specifically collected for this thesis: a corpus of Wikipedia articles about mobile phones and a corpus of mobile phone reviews published on the Epinions.com website. Parts of these corpora were manually annotated with coreference relations, mobile phone features and relations between mentions of the phone and its features. The annotation is used to develop a coreference resolution module as well as a machine learning-based relation extractor. Rule-based methods for identification of coreference chains describing the phone are designed and thoroughly evaluated against the annotated gold standard. Machine learning is used to find links between mentions of the phone (identified by coreference resolution) and phone features. It determines whether some phone feature belong to the phone mentioned in the same sentence or not. In order to find the best rankings, this thesis investigates several settings. One of the hypotheses tested here is that the relatively low results of the proposed baseline are caused by noise introduced by sentences which are not directly related to the phone and phone feature. To test this hypothesis, only sentences which contained mentions of the mobile phone and a phone feature linked to it were processed to produce rankings of the phones features. Selection of the relevant sentences is based on the results of coreference resolution and relation extraction. Another hypothesis is that opinionated sentences are a good source for ranking the phone features. In order to investigate this, a sentiment classification system is also employed to distinguish between features mentioned in positive and negative contexts. The detailed evaluation and error analysis of the methods proposed form an important part of this research and ensure that the results provided in this thesis are reliable.

The Data Web has undergone a tremendous growth period. It currently consists of more then 3300 publicly available knowledge bases describing millions of resources from various domains, such as life sciences, government or geography, with over 89 billion facts. In the same way, the Document Web grew to the state where approximately 4.55 billion websites exist, 300 million photos are uploaded on Facebook as well as 3.5 billion Google searches are performed on average every day. However, there is a gap between the Document Web and the Data Web, since for example knowledge bases available on the Data Web are most commonly extracted from structured or semi-structured sources, but the majority of information available on the Web is contained in unstructured sources such as news articles, blog post, photos, forum discussions, etc. As a result, data on the Data Web not only misses a significant fragment of information but also suffers from a lack of actuality since typical extraction methods are time-consuming and can only be carried out periodically. Furthermore, provenance information is rarely taken into consideration and therefore gets lost in the transformation process. In addition, users are accustomed to entering keyword queries to satisfy their information needs. With the availability of machine-readable knowledge bases, lay users could be empowered to issue more specific questions and get more precise answers. In this thesis, we address the problem of Relation Extraction, one of the key challenges pertaining to closing the gap between the Document Web and the Data Web by four means. First, we present a distant supervision approach that allows finding multilingual natural language representations of formal relations already contained in the Data Web. We use these natural language representations to find sentences on the Document Web that contain unseen instances of this relation between two entities. Second, we address the problem of data actuality by presenting a real-time data stream RDF extraction framework and utilize this framework to extract RDF from RSS news feeds. Third, we present a novel fact validation algorithm, based on natural language representations, able to not only verify or falsify a given triple, but also to find trustworthy sources for it on the Web and estimating a time scope in which the triple holds true. The features used by this algorithm to determine if a website is indeed trustworthy are used as provenance information and therewith help to create metadata for facts in the Data Web. Finally, we present a question answering system that uses the natural language representations to map natural language question to formal SPARQL queries, allowing lay users to make use of the large amounts of data available on the Data Web to satisfy their information need.

Textmining beschäftigt sich mit der automatisierten Annotierung von Texten und der Extraktion einzelner Informationen aus Texten, die dann für die Weiterverarbeitung zur Verfügung stehen. Texte können dabei kurze Zusammenfassungen oder komplette Artikel sein, zum Beispiel Webseiten und wissenschaftliche Artikel, umfassen aber auch textuelle Einträge in sonst strukturierten Datenbanken. Diese Dissertationsschrift bespricht zwei wesentliche Themen des biomedizinischen Textmining: die Extraktion von Zusammenhängen zwischen biologischen Entitäten ---das Hauptaugenmerk liegt dabei auf der Erkennung von Protein-Protein-Interaktionen---, und einen notwendigen Vorverarbeitungsschritt, die Erkennung von Proteinnamen. Diese Schrift beschreibt Ziele, Herausforderungen, sowie typische Herangehensweisen für alle wesentlichen Komponenten des biomedizinischen Textmining. Wir stellen eigene Methoden zur Erkennung von Proteinnamen sowie der Extraktion von Protein-Protein-Interaktionen vor. Zwei eigene Verfahren zur Erkennung von Proteinnamen werden besprochen, eines basierend auf einem Klassifikationsproblem, das andere basierend auf Suche in Wörterbüchern. Für die Extraktion von Interaktionen entwickeln wir eine Methode zur automatischen Annotierung großer Mengen von Text im Bezug auf Relationen; diese Annotationen werden dann zur Mustererkennung verwendet, um anschließend die gefundenen Muster auf neuen Text anwenden zu können. Um Muster zu erkennen, berechnen wir Ähnlichkeiten zwischen zuvor gefundenen Sätzen, die denselben Typ von Relation/Interaktion beschreiben. Diese Ähnlichkeiten speichern wir als sogenannte `consensus patterns''. Wir entwickeln eine Alignmentstrategie, die mehrschichtige Annotationen pro Position im Muster erlaubt. In Versuchen auf bekannten Benchmarks zeigen wir empirisch, dass unser vollautomatisches Verfahren Resultate erzielt, die vergleichbar sind mit existierenden Methoden, welche umfangreiche Eingriffe von Experten voraussetzen. / Text mining deals with the automated annotation of texts and the extraction of facts from textual data for subsequent analysis. Such texts range from short articles and abstracts to large documents, for instance web pages and scientific articles, but also include textual descriptions in otherwise structured databases. This thesis focuses on two key problems in biomedical text mining: relationship extraction from biomedical abstracts ---in particular, protein--protein interactions---, and a pre-requisite step, named entity recognition ---again focusing on proteins. This thesis presents goals, challenges, and typical approaches for each of the main building blocks in biomedical text mining. We present out own approaches for named entity recognition of proteins and relationship extraction of protein-protein interactions. For the first, we describe two methods, one set up as a classification task, the other based on dictionary-matching. For relationship extraction, we develop a methodology to automatically annotate large amounts of unlabeled data for relations, and make use of such annotations in a pattern matching strategy. This strategy first extracts similarities between sentences that describe relations, storing them as consensus patterns. We develop a sentence alignment approach that introduces multi-layer alignment, making use of multiple annotations per word. For the task of extracting protein-protein interactions, empirical results show that our methodology performs comparable to existing approaches that require a large amount of human intervention, either for annotation of data or creation of models.

<p>This thesis describes and introduces novel methods for knowledge discovery and presents a software system that is able to extract information from biomedical literature, review interesting connections between various biomedical concepts and in so doing, generates new hypotheses. The experimental results obtained by using methods described in this thesis, are compared to currently published results obtained by other methods and a number of case studies are described. This thesis shows how the technology&nbsp / resented can be integrated with the researchers&rsquo / own knowledge, experimentation and observations for optimal progression of scientific research.</p>

This thesis aims to develop a Relation Extraction algorithm to extract knowledge out of automotive data. While most approaches to Relation Extraction are only evaluated on newspaper data dealing with general relations from the business world their applicability to other data sets is not well studied. Part I of this thesis deals with theoretical foundations of Information Extraction algorithms. Text mining cannot be seen as the simple application of data mining methods to textual data. Instead, sophisticated methods have to be employed to accurately extract knowledge from text which then can be mined using statistical methods from the field of data mining. Information Extraction itself can be divided into two subtasks: Entity Detection and Relation Extraction. The detection of entities is very domain-dependent due to terminology, abbreviations and general language use within the given domain. Thus, this task has to be solved for each domain employing thesauri or another type of lexicon. Supervised approaches to Named Entity Recognition will not achieve reasonable results unless they have been trained for the given type of data. The task of Relation Extraction can be basically approached by pattern-based and kernel-based algorithms. The latter achieve state-of-the-art results on newspaper data and point out the importance of linguistic features. In order to analyze relations contained in textual data, syntactic features like part-of-speech tags and syntactic parses are essential. Chapter 4 presents machine learning approaches and linguistic foundations being essential for syntactic annotation of textual data and Relation Extraction. Chapter 6 analyzes the performance of state-of-the-art algorithms of POS tagging, syntactic parsing and Relation Extraction on automotive data. The findings are: supervised methods trained on newspaper corpora do not achieve accurate results when being applied on automotive data. This is grounded in various reasons. Besides low-quality text, the nature of automotive relations states the main challenge. Automotive relation types of interest (e. g. component – symptom) are rather arbitrary compared to well-studied relation types like is-a or is-head-of. In order to achieve acceptable results, algorithms have to be trained directly on this kind of data. As the manual annotation of data for each language and data type is too costly and inflexible, unsupervised methods are the ones to rely on. Part II deals with the development of dedicated algorithms for all three essential tasks. Unsupervised POS tagging (Chapter 7) is a well-studied task and algorithms achieving accurate tagging exist. All of them do not disambiguate high frequency words, only out-of-lexicon words are disambiguated. Most high frequency words bear syntactic information and thus, it is very important to differentiate between their different functions. Especially domain languages contain ambiguous and high frequent words bearing semantic information (e. g. pump). In order to improve POS tagging, an algorithm for disambiguation is developed and used to enhance an existing state-of-the-art tagger. This approach is based on context clustering which is used to detect a word type’s different syntactic functions. Evaluation shows that tagging accuracy is raised significantly. An approach to unsupervised syntactic parsing (Chapter 8) is developed in order to suffice the requirements of Relation Extraction. These requirements include high precision results on nominal and prepositional phrases as they contain the entities being relevant for Relation Extraction. Furthermore, accurate shallow parsing is more desirable than deep binary parsing as it facilitates Relation Extraction more than deep parsing. Endocentric and exocentric constructions can be distinguished and improve proper phrase labeling. unsuParse is based on preferred positions of word types within phrases to detect phrase candidates. Iterating the detection of simple phrases successively induces deeper structures. The proposed algorithm fulfills all demanded criteria and achieves competitive results on standard evaluation setups. Syntactic Relation Extraction (Chapter 9) is an approach exploiting syntactic statistics and text characteristics to extract relations between previously annotated entities. The approach is based on entity distributions given in a corpus and thus, provides a possibility to extend text mining processes to new data in an unsupervised manner. Evaluation on two different languages and two different text types of the automotive domain shows that it achieves accurate results on repair order data. Results are less accurate on internet data, but the task of sentiment analysis and extraction of the opinion target can be mastered. Thus, the incorporation of internet data is possible and important as it provides useful insight into the customer\'s thoughts. To conclude, this thesis presents a complete unsupervised workflow for Relation Extraction – except for the highly domain-dependent Entity Detection task – improving performance of each of the involved subtasks compared to state-of-the-art approaches. Furthermore, this work applies Natural Language Processing methods and Relation Extraction approaches to real world data unveiling challenges that do not occur in high quality newspaper corpora.

<p>This thesis describes and introduces novel methods for knowledge discovery and presents a software system that is able to extract information from biomedical literature, review interesting connections between various biomedical concepts and in so doing, generates new hypotheses. The experimental results obtained by using methods described in this thesis, are compared to currently published results obtained by other methods and a number of case studies are described. This thesis shows how the technology&nbsp / resented can be integrated with the researchers&rsquo / own knowledge, experimentation and observations for optimal progression of scientific research.</p>

The overwhelming amount of online news presents a challenge called news information overload. To mitigate this challenge we propose a system to generate a causal network of news topics. To extract this information from distributed news sources, a system called Forest was developed. Forest retrieves documents that potentially contain causal information regarding a news topic. The documents are processed at a sentence level to extract causal relations and news topic references, these are the phases used to refer to a news topic. Forest uses a machine learning approach to classify causal sentences, and then renders the potential cause and effect of the sentences. The potential cause and effect are then classified as news topic references, these are the phrases used to refer to a news topics, such as “The World Cup” or “The Financial Meltdown”. Both classifiers use an algorithm developed within our working group, the algorithm performs better than several well known classification algorithms for the aforementioned tasks. In our evaluations we found that participants consider causal information useful to understand the news, and that while we can not extract causal information for all news topics, it is highly likely that we can extract causal relation for the most popular news topics. To evaluate the accuracy of the extractions made by Forest, we completed a user survey. We found that by providing the top ranked results, we obtained a high accuracy in extracting causal relations between news topics.

L’abondance de textes dans le domaine biomédical nécessite le recours à des méthodes de traitement automatique pour améliorer la recherche d’informations précises. L’extraction d’information (EI) vise précisément à extraire de l’information pertinente à partir de données non-structurées. Une grande partie des méthodes dans ce domaine se concentre sur les approches d’apprentissage automatique, en ayant recours à des traitements linguistiques profonds. L’analyse syntaxique joue notamment un rôle important, en fournissant une analyse précise des relations entre les éléments de la phrase.Cette thèse étudie le rôle de l’analyse syntaxique en dépendances dans le cadre d’applications d’EI dans le domaine biomédical. Elle comprend l’évaluation de différents analyseurs ainsi qu’une analyse détaillée des erreurs. Une fois l’analyseur le plus adapté sélectionné, les différentes étapes de traitement linguistique pour atteindre une EI de haute qualité, fondée sur la syntaxe, sont abordés : ces traitements incluent des étapes de pré-traitement (segmentation en mots) et des traitements linguistiques de plus haut niveau (lié à la sémantique et à l’analyse de la coréférence). Cette thèse explore également la manière dont les différents niveaux de traitement linguistique peuvent être représentés puis exploités par l’algorithme d’apprentissage. Enfin, partant du constat que le domaine biomédical est en fait extrêmement diversifié, cette thèse explore l’adaptation des techniques à différents sous-domaines, en utilisant des connaissances et des ressources déjà existantes. Les méthodes et les approches décrites sont explorées en utilisant deux corpus biomédicaux différents, montrant comment les résultats d’IE sont utilisés dans des tâches concrètes. / The abundance of biomedical information expressed in natural language has resulted in the need for methods to process this information automatically. In the field of Natural Language Processing (NLP), Information Extraction (IE) focuses on the extraction of relevant information from unstructured data in natural language. A great deal of IE methods today focus on Machine Learning (ML) approaches that rely on deep linguistic processing in order to capture the complex information contained in biomedical texts. In particular, syntactic analysis and parsing have played an important role in IE, by helping capture how words in a sentence are related. This thesis examines how dependency parsing can be used to facilitate IE. It focuses on a task-based approach to dependency parsing evaluation and parser selection, including a detailed error analysis. In order to achieve a high quality of syntax-based IE, different stages of linguistic processing are addressed, including both pre-processing steps (such as tokenization) and the use of complementary linguistic processing (such as the use of semantics and coreference analysis). This thesis also explores how the different levels of linguistics processing can be represented for use within an ML-based IE algorithm, and how the interface between these two is of great importance. Finally, biomedical data is very heterogeneous, encompassing different subdomains and genres. This thesis explores how subdomain-adaptationcan be achieved by using already existing subdomain knowledge and resources. The methods and approaches described are explored using two different biomedical corpora, demonstrating how the IE results are used in real-life tasks.

Philosophiae Doctor - PhD / This thesis describes and introduces novel methods for knowledge discovery and presents a software system that is able to extract information from biomedical literature, review interesting connections between various biomedical concepts and in so doing, generates new hypotheses. The experimental results obtained by using methods described in this thesis, are compared to currently published results obtained by other methods and a number of case studies are described. This thesis shows how the technology, resented can be integrated with the researchers own knowledge, experimentation and observations for optimal progression of scientific research. / South Africa