‘Tecnologica cosa’: Modeling Storyteller Personalities in Boccaccio’s ‘Decameron’

Abstract

We explore Boccaccio's Decameron to see how digital humanities tools can be used for tasks that have limited data in a language no longer in contemporary use: medieval Italian. We focus our analysis on the question: Do the different storytellers in the text exhibit distinct personalities? To answer this question, we curate and release a dataset based on the authoritative edition of the text. We use supervised classification methods to predict storytellers based on the stories they tell, confirming the difficulty of the task, and demonstrate that topic modeling can extract thematic storyteller "profiles."

Four Years of FAccT: A Reflexive, Mixed-Methods Analysis of Research Contributions, Shortcomings, and Future Prospects

“…We therefore looked to prior successful small-text topic analyses to inform our training procedure. Following recently published work [20], we use a Python wrapper of the MALLET library to train our model [5,59]. MALLET, unlike the more-popular Python-based gensim library, uses Gibbs sampling [32] for the LDA algorithm's underlying sampling method.…”

“…MALLET, unlike the more-popular Python-based gensim library, uses Gibbs sampling [32] for the LDA algorithm's underlying sampling method. Gibbs sampling is an exact Markov chain Monte Carlo technique [15], which Cooper et al [20] notes has better performance for small-text corpora than inexact, variational-inference LDA implementations [45]. For the documents submitted to LDA in training, we chunk the FAccT papers into contiguous segments of 200 words, following the intuition concerning hyperparameter tuning for Gibbs-sampling-based LDA described in the well-cited paper by Griffiths and Steyvers [35].…”

“…Choosing the number of topics 𝑘 for the model to learn requires some domain expertise and some degree of human intervention [17,20,35]. We began our experiments with 𝑘 = 20, and tried larger and smaller 𝑘, guided by general advice in Griffiths and Steyvers [35], and the results reported reflect those that worked the best when confirming the output topics (𝑘 = 22).…”

“…We began our experiments with 𝑘 = 20, and tried larger and smaller 𝑘, guided by general advice in Griffiths and Steyvers [35], and the results reported reflect those that worked the best when confirming the output topics (𝑘 = 22). From performing this process for the 5 trained models (𝑘 = 19,20,21,22,23), we selected the model with 𝑘 = 22 topics for our analysis. A selection of topic labels, with a subset of topic-specific words, can be found in Table 5.…”