MuLX-QA: Classifying Multi-Labels and Extracting Rationale Spans in Social Media Posts

While social media platforms play an important role in our daily lives in obtaining the latest news and trends from across the globe, they are known to be prone to widespread proliferation of harmful information in different forms leading to misconceptions among the masses. Accordingly, several prior works have attempted to tag social media posts with labels/classes reflecting their veracity, sentiments, hate content, etc. However, in order to have a convincing impact, it is important to additionally extract the post snippets on which the labelling decision is based. We call such a post snippet as the ‘rationale’. These rationales significantly improve human trust and debuggability of the predictions, especially when detecting misinformation or stigmas from social media posts. These rationale spans or snippets are also helpful in post-classification social analysis, such as for finding out the target communities in hate-speech, or for understanding the arguments or concerns against the intake of vaccines. Also it is observed that a post may express multiple notions of misinformation, hate, sentiment, etc. Thus, the task of determining (one or multiple) labels for a given piece of text, along with the text snippets explaining the rationale behind each of the identified labels is a challenging multi-label, multi-rationale classification task, which is still nascent in the literature.

While transformer-based encoder-decoder generative models such as BART and T5 are well-suited for the task, in this work we show how a relatively simpler encoder-only discriminative question-answering (QA) model can be effectively trained using simple template-based questions to accomplish the task. We thus propose MuLX-QA and demonstrate its utility in producing (label, rationale span) pairs in two different settings: multi-class (on the HateXplain dataset related to hate speech on social media), and multi-label (on the CAVES dataset related to COVID-19 anti-vaccine concerns). MuLX-QA outperforms heavier generative models in both settings. We also demonstrate the relative advantage of our proposed model MuLX-QA over strong baselines when trained with limited data. We perform several ablation studies, and experiments to better understand the effect of training MuLX-QA with different question prompts, and draw interesting inferences. Additionally, we show that MuLX-QA is effective on social media posts in resource-poor non-English languages as well. Finally, we perform a qualitative analysis of our model predictions and compare them with those of our strongest baseline.