How Explainable is Automatic Hand Identification? Case Studies and Perspectives
Abstract. Starting with some case studies, in this contribution we address the issue of explainability and interpretability of results from automatic handwriting identification, in a way that domain experts can understand and validate.
In recent years, the adoption of computational methods to automate the analysis of medieval and modern manuscripts has included the application of multiple artificial intelligence (deep learning) algorithms to, e.g., identify hands, perform dating, and extract entities of interest. Despite high performance in some limited domains, two critical drawbacks have been found: first, the lack of structured data in the paleographical domain in the form of fully annotated manuscripts, which are costly to obtain and impossible to automate via, e.g., the use of large language models. Second, the so-called “black box” problem, that consists of representing solutions obtained by complex algorithms in a way that domain experts can understand and validate.
We present here our current research to tackle these problems, with a focus on the interpretability component, such as the ongoing research on the ms. Vat. lat. 653, written by four different scribes. We investigate a wide array of explainability techniques to benchmark their usefulness. First, we propose a “post-hoc” solution based on the LIME (Local Interpretable Model-agnostic Explanations), specifically tailored for the HI task, which aims at highlighting the impact that a portion of the manuscript has on the assignment to the scribe who wrote it. We then investigate the use of data attribution methods to connect predictions with examples shown during the training of the model, to understand which annotations provided by the paleographers were beneficial (or detrimental) to a given hand assignment.
Overall, the results are encouraging; nonetheless, we observe via visual inspection that existing explainability methods have limited interpretability, as they do not correspond to semantically rich features such as those used by paleographers for the hand identification task. As future work, we plan to develop fully interpretable and controllable AI models, whose training and inference can be guided by the interaction between scholars and the explanation system, to develop a system that could learn higher-level features through a systematic feedback loop between the paleographer and the system.