Entity tagging and relation extraction from historical texts using deep learning

The process of assigning metadata to spans of text in order to encode meaning, or “marking up”, has been central to digital humanities practice since before the field had a name. Thousands of person-hours have been devoted to this task, either performing the process entirely manually, or partially assisted by NLP software. Since the 2010s, deep learning has profoundly impacted NLP, greatly expanding the range of language tasks which can be performed automatically or semi-automatically. Transformers in particular have improved the performance of tasks such as Named Entity Recognition (NER) and Relation Extraction (RE) to the point where it is plausible that markup tasks previously performed by skilled research assistants can, with appropriate training data, be automated or partially automated. The Digital Humanities Institute in Sheffield has been involved in many large scale markup projects over the past twenty-five years and amassed a huge tranche of potential training data for these tasks. To what extent can deep learning models accurately learn and replicate the work of the researchers who first tagged this material in the 2000s and 2010s? This walkthrough will instruct participants in how to set up and train a transformer model to mark up defendants, crimes and verdicts in text from the Old Bailey Proceedings and correctly identify the relationships between them. It will also explore how to deduce the optimum training strategy for any similar markup task. For example, the number of ground truth human examples required to produce good results, or shortcuts such as using a partially trained transformer to assist a semi-automated workflow. Ultimately this technology should allow relevant digital humanities projects to be delivered at lower cost and allocate more time and resources to data analysis rather than markup, promoting research excellence.A recording of this session is available on YouTube: https://youtu.be/evmHy5czdpw