Biographical information extraction: A language-agnostic methodology for datasets and models

Abstract

Information extraction (IE) refers to the task of detecting and linking information contained in written texts. While it includes various subtasks, relation extraction (RE) is used to link two entities in a text via a common relation. RE can therefore be used to build linked databases of knowledge across a wide area of topics. Today, the task of RE is treated as a supervised machine learning (ML) task, where a model is trained using a specific architecture and a specific annotated dataset. These specific datasets typically aim to represent common patterns that the model is to learn, albeit at the cost of manual annotation, which can be costly and time-consuming. In addition, due to the nature of the training process, the models can be sensitive to a specific genre or topic, and are generally monolingual. It therefore stands to reason, that certain genres and topics have better models, as they are treated with a higher priority due to financial interests for instance. This in turn leads to RE models not being available to every area of research, leaving incomplete linked databases of knowledge. For instance, if the birthplace of a person is not correctly extracted, the place and the person can not be linked correctly, therefore not leaving linked databases incomplete. To address this problem, this thesis explores aspects of RE that could be adapted in ways which require little human effort, therefore making RE models more widely available. The first aspect is the annotated data. During the course of this thesis, Wikipedia and its subsidiaries are used as sources to automatically annotate sentences for RE. The dataset, which is aimed towards digital humanities (DH) and historical research, is automatically compiled by aligning sentences from Wikipedia articles with matching structured data from sources including Pantheon and Wikidata. By exploiting the structure of Wikipedia articles and robust named entity recognition (NER), information is matched with relatively high precision in order to compile annotated relation pairs for ten different relations that are important in the DH domain: birthdate, birthplace, deathdate, deathplace, occupation, parent, educated, child, sibling and other (all other relations). Furthermore, the effectiveness of the dataset is demonstrated by training a state-of-the-art neural model to classify relation pairs. For its evaluation, a manually annotated gold standard set is used. An investigation of the necessary adaptations to recreate the automatic process in a multilingual setting is also undertaken, looking specifically at English and German, for which similar neural models are trained and evaluated on a gold standard dataset. While the process is aimed here at training neural models for RE within the domain of digital humanities and history, it may be transferable to other domains.