RePALM: Popular Quote Tweet Generation via Auto-Response Augmentation

A quote tweet allows users to share others' content and add their comments. To help users write a quote tweet with better public engagement, we study the popular quote tweet generation task. It aims to generate quote tweets with higher popularity reflected by more likes, replies, and retweets. Despite the exceptional language generation capabilities of large language models (LLMs), limited work has examined how LLMs can learn the popularity of text to engage the public better. Consequently, we propose a novel Response-augmented Popularity-Aligned Language Model (RePALM) to align language generation to popularity by leveraging readers' insights from augmented auto-responses. Here, we employ the Proximal Policy Optimization framework with a dual-reward mechanism to jointly explore the quote tweet's popularity and consistency with the auto-responses. For experiments, we gathered two datasets of quote tweets with external links and others' tweets. Extensive results show the superiority of RePALM over advanced language models without response augmentation.