Individual Submission Summary
Share...

Direct link:

Troubles in Text: Finetuning NLP to Analyze Rationalization of Rights Violations

Sun, October 3, 10:00 to 11:30am PDT (10:00 to 11:30am PDT), TBA

Abstract

Can innovations in natural language processing (NLP) be adapted to improve language modeling of complex political concepts? Computational uses of “bags of words” (BOW) allowed political scientists to systematically analyze large text collections. However, BOWs do not capture a word’s contextual meaning and are therefore ill-equipped to recognize complex, “know-when-I-see-it” (kwIsi) concepts. NLP’s pre-train/finetune paradigm—which dramatically improved computer science language modeling—uses contextualized word embeddings and is better suited to capture kwIsi concepts. This article models the pre-train/finetune approach for political science and evaluates RoBERTa (Liuet al.2019)’s performance modeling a kwIsi concept: archived British government rationalizations for internment without trial (1971-75). RoBERTa outperforms conventional BOW-based methods at classifying an internment rationalization but faces challenges identifying rare-case rationalizations in imperfect text corpora. We highlight promises for NLP applications in political science, when grounded in qualitative methods, and identify challenges to consider when modeling kwIsi concepts in political texts.

Authors