Daidalos: NER for Literary Studies on Latin and Ancient Greek Texts

Daidalos: NER for Literary Studies on Latin and Ancient Greek Texts Nomina Omina: Detecting and Preserving Ancient Greek and Latin Proper Names in the Age of Artificial Intelligence, Leipzig, 27/06/2024 Dr. Andrea Beyer (Humboldt-Universität zu Berlin) Daidalos 01 NER in Research: Standalone Method 02 NER in Research: Part of a Pipeline 03 NER in Teaching 04 Named Entity Recognition for Literary Studies on Latin and Ancient Greek Texts 01 | Daidalos Project Infrastructure Goals Why Call a Project “Daidalos”? We … ─develop an NLP infrastructure ─that will enable researchers in Classical Philology and related disciplines ─to apply various methods of natural language processing ─which are uncommon in the German speaking philological community. I was the most famous inventor, craftsman, and builder in antiquity –forget my human failures. daidalos-projekt.de Daidalos Platform Menu: NLP-Tools ☑Select: language, author, work, text passage ☑Run ☑Choose between NLP methods NER, POS, Sentiment Analysis daidalos-projekt.de Infrastructure Multiple NLP methods and corpora, adjustable settings, pipelines for literary research questions, Identity & Access Management Community of Practice OA-Publication with research tandems, learning opportunities (Jupyter Notebooks, H5P), data bases on tools and literature, workshops Interpretable AI Transparency & sustainability by using model cards, data sheets, and well documented evaluations of methods Goals 02 | NER in Research: Standalone Method Example Tagger: Quality & Applications Challenges & Solutions Example Tagger: Quality & Applications Latin Ancient Greek Model Name la_core_web_lg UGARIT/flair_grc_bert_ner Publication Burns 2023 Yousef et al. 2023 NLP Software spaCy Flair NLP Architecture floret vectors Transition-based Parser BERT (Transformer)vectors Long Short-Term Memory network Conditional Random Field Training Data Caesar, Ovid, Pliny (Elder & Younger) Homer, Herodotus, Athenaeus Tagset persons, locations persons, locations, peoples Which NER Tagger Should You Use? Model Cards & Datasheets offer an overview How Do You Learn to Use NER? Curated Jupyter Notebooks provide an introduction Why Should You Learn to Use NER? Understanding NER is part of improving one‘s own Digital Literacies Teaching is About What, How, and Why Model Cards … …accompany the models and provide handy information …can be Markdown files with additional metadata …are essential for discoverability, reproducibility, and sharing But model cards are difficult … …to understand by average researchers who lack the necessary digital literacies …to compare with each other for selecting the most suitable tagger Model cards should describe … …the model, its intended use, potential limitations, including biases and ethical considerations, the data, selection for training and evaluation, possible limitations, and recommendations, if necessary Model Card https://anonymous.4open.science/r/seflag-DC3B/documentation/model_cards/latincy.md Datasheets … …offer question-driven information about the dataset of a model …include questions on possible sensitive data But datasheets might contain too much information that is not structured enough for unexperienced users / researchers. Datasheet https://anonymous.4open.science/r/seflag-DC3B/documentation/datasheet_latin.md (excerpt: only first paragraph) Jupyter Notebooks as Interactive Worksheets ─Jupyter Notebooks are files that contain interactive worksheets ─Code can be supplemented with a. Text b. Coloured boxes c. Table of contents d. Integration of graphics or videos e. … ─Aim: acquisition of new learning content, more in-depth study or repetition, easy access to digital methods But working with Jupyter Notebooks is much more demanding than it may seem at first … Overview Short method definition Embedding in research topic Approach Expected result Level 1 AI Literacy Understand the method Fully guided Use given example Challenges Using Jupyter Notebooks Generalisation unclear (e.g. any text) Technical vocabulary (e.g. library) Running code and dealing with potential error messages (software dependencies) Challenges Connect explanation with code snippets Comprehend technical outputs Understand and interpret results (e.g. result accuracy for each entity) Challenges HTML Dealing with incorrect results Understanding limits and opportunities of this method