Session 7: KI in den Geschichtswissenschaften I – Verfahren des maschinelles Lernens
|Datum & Zeit:
15:20 – 16:20 Uhr
|Auditorium des Jacob-und-Wilhelm-Grimm-Zentrums
Impact of AI: Gamechanger for Image Classification in Historical Research?
Michela Vignoli1, Doris Gruber2, Rainer Simon1, Axel Weißenfeld1
1AIT Austrian Institute of Technology, Österreich; 2Österreichische Akademie der Wissenschaften, Österreich
AI opens new possibilities for processing and analysing large, heterogeneous historical data corpora in a semi-automated way. The Ottoman Nature in Travelogues (ONiT) project develops an interdisciplinary methodological framework for an AI-driven analysis of text–image relations in digitised printed material. The goal is to provide a workflow and tools that support the semi-automatic detection and classification of “nature” representations in both text and images from travel accounts on the Ottoman empire. Our leading research questions are, what role representations of “nature” played in the reports, whether and, if so, why differences occurred in diachronic and synchronic perspectives, and how the texts and images relate to each other. Based on the results and learnings, ONiT will discuss to what extent computational, quantitative methods enable us to answer these questions, and in which way they affect our underlying epistemology stemming from more traditional “analogue” methods. Our experiences so far confirm that the interdisciplinary collaboration between historians and AI developers has a notable impact on the methods to be applied. One of our main learnings is the necessity to develop an understanding for distinct visual features as opposed to representations of “nature” that require interpretation. For example, a drawing of a natural landscape will most likely contain representation of plants even though they are not clearly depicted. But is the AI algorithm capable of generalising this from the provided training data? Such considerations are relevant to discuss how AI technologies affect hermeneutics and conclusions related to the analysed data corpus in history research.
Technological change and sentiment in German parliamentary speeches (1867-1932)
1Humboldt-Universität zu Berlin; 2Max-Planck-Institut für Rechtsgeschichte und Rechtstheorie, Frankfurt am Main; 3Centrum für Europäische Politik (cep), Berlin
In recent years, parliamentary debate transcripts have been increasingly used for sentiment analysis, albeit, so far, only in the English-language context. This paper analyses the digital corpus of German parliamentary speeches from 1867 to 1932 and compares different types of sentiment analysis for their usage in historical scholarship. Dictionary methods are an exceptionally transparent way to quantify sentiment. Still, they also suffer from problems: They typically (albeit not in all variants) miss negators and adversative conjunctions, cannot understand satire and sarcasm and might be ill-suited to earlier periods, as they are usually derived from contemporary language. Political speeches are more complex than standardised product reviews and social media posts that are usually quantified by sentiment scores, which makes it more difficult to transfer this NLP method to the historical sciences. Especially speeches about ‘innovation’ are often long, cover diverse topics, and include multiple targets of subjective language. Therefore, it might be necessary to turn to more sophisticated methods, such as linear classifiers trained on a bag-of-words text representation or neural networks trained using a transformer word embedding model (BERT). The most robust procedure might entail manually creating smaller but more controlled dictionaries with the help of domain knowledge, which can then be used as a list of ‘seed words’ to train a sentiment algorithm for the specific historical application. Besides enabling a methodological comparison of different approaches to sentiment analysis, the empirical results of this paper can be compared with findings from the qualitative literature.