The study of classroom discourse is essential for enhancing child development and educational outcomes in academic settings. Prior research has focused on the annotation of conversational talk-turns within the classroom, offering a statistical analysis of the various types of discourse prevalent in these environments. In this work, we explore the generalizability and transferability of text classifiers trained to predict these discourse codes across educational domains. We examine two distinct English-language classroom datasets from the domains: literacy and math. Our results show that models exhibit high accuracy and generalizability when the training and test datasets originate from the same or similar domains. In situations where limited training data is available in new domains, few shot and zero shot exhibit more resiliency and aren’t as effected as their supervised counterparts. We also observe that accompanying each talk turn with dialog-level context improves the accuracy of the generative models. We conclude by offering suggestions on how to enhance the generalization of these methods to novel domains, proposing directions for future studies to investigate new methods for boosting the model adaptability across domains.

Content moderation is essential in preventing the spread of harmful content on the Internet. However, there are instances where moderation fails and it is important to understand when and why that happens. Workflows that aim to uncover a system’s weakness typically use clustering of the data points’ embeddings to group errors together. In this paper, we evaluate the K-Means clustering of four text representations for the task of offensive language detection in English and Levantine Arabic. We find Sentence-BERT (SBERT) embeddings give the most human-interpretable clustering for English errors and the grouping is mainly based on the targeted group in the text. Meanwhile, SBERT embeddings of Large Language Model (LLM)-generated linguistic features give the most interpretable clustering for Arabic errors.

In recent years, gray social media platforms, those with a loose moderation policy on cyberbullying, have been attracting more users. Recently, data collected from these types of platforms have been used to pre-train word embeddings (social-media-based), yet these word embeddings have not been investigated for social NLP related tasks. In this paper, we carried out a comparative study between social-media-based and non-social-media-based word embeddings on two social NLP tasks: Detecting cyberbullying and Measuring social bias. Our results show that using social-media-based word embeddings as input features, rather than non-social-media-based embeddings, leads to better cyberbullying detection performance. We also show that some word embeddings are more useful than others for categorizing offensive words. However, we do not find strong evidence that certain word embeddings will necessarily work best when identifying certain categories of cyberbullying within our datasets. Finally, We show even though most of the state-of-the-art bias metrics ranked social-media-based word embeddings as the most socially biased, these results remain inconclusive and further research is required.