bert-quote-attribution and Grimbert

@GitLab

During my master's thesis, I wanted to extract conversational character networks from novels. To do so, I had to automatically detect quote speakers in novels: that is, perform quote attribution. Since addressee attribution was a much harder task, I did not tackle it at the time.

That was the start of the bert-quote-attribution project. at the time, BERT (1) was very new, and I used it to generate representations from quotes and potential candidate speakers. I then trained a ranker based on those features, that would output a score between 0 and 1 to decide whether or not a candidate was a quote speaker. The ranker was trained on a dataset of 3 classical novels introduced by Muzny et al., 2017 (2). See my master's thesis for more details. Documentation is available at https://aethor.gitlab.io/bert-quote-attribution.

At the time, visualizing attention was all the rage, so the project even includes an attention visualizer:

@GitHub

Grimbert is a new and improve version of my previous bert-quote-attribution project, and is used in Renard. Compared to its previous iteration, Grimbert is installable using pip, the training loop uses huggingface instead of being manually implemented which allows more parameters, the training can be monitored by sacred, and it supports the PDNC (3) dataset as well as the original one from Muzny et al., 2017 (2).