Lunch at 12:30, talk at 1pm, in 148 Fitzpatrick

Title: Making Generic Frames Visible: Building Neural Network Language Models to Classify Frames

Abstract: Supervised machine learning has become a helpful tool for communication scholars to automate news articles’ frame classification. Recent developments in machine learning allow us to analyze texts by considering words and sentences’ context. Yet, these new developments have not been tested to classify generic frames in news articles. We advance prior research by evaluating the classification performance of several neural network language models (i.e., BERT, FastText, and ELMO) and comparing these models to baselines and non-contextual language models (e.g., Naive Bayes, TF-IDF). We tested these models using a manually coded dataset of Chilean news articles, and our results show differences in performance across four generic frames. While contextual language models detected more conflict frames correctly than non-contextualized language models, the latter outperformed the former when classifying economic and human-interest frames. These results shed light on the benefits and limitations of using neural network language models to identify and code frames in news articles.

Bio: Professor Gómez-Zará’s research focuses on how social computational systems help people organize and collaborate. His work has been at the forefront of computational social science, human-computer interaction, and network science.