Causal learning in question quality improvement

To improve the quality of questions asked in Community-based questions answering forums, we create a new dataset from the website, Stack Overflow, which contains three components: (1) context: the text features of questions, (2) treatment: categories of revision suggestions and (3) outcome: the measure of question quality (e.g., the number of questions, upvotes or clicks). This dataset helps researchers develop causal inference models towards solving two problems: (i) estimating the causal effects of aforementioned treatments on the outcome and (ii) finding the optimal treatment for the questions. Empirically, we performed experiments with three state-of-the-art causal effect estimation methods on the contributed dataset. In particular, we evaluated the optimal treatments recommended by the these approaches by comparing them with the ground truth labels – treatments (suggestions) provided by experts.