A Two Stage Adaptation Framework for Frame Detection via Prompt Learning

Proceedings of the 29th International Conference on Computational Linguistics, pages 2968–2978 October 12–17, 2022. 2968 A Two Stage Adaptation Framework for Frame Detection via Prompt Learning Xinyi Mou1, Zhongyu Wei1,2∗, Changjian Jiang3, Jiajie Peng4 1School of Data Science, Fudan University, China 2Research Institute of Intelligent and Complex Systems, Fudan University, China 3School of International Relations & Public Affairs, Fudan University, China 4School of Computer Science, Northwestern Polytechnical University, China {xymou20,zywei,Changjian}@fudan.edu.cn jiajiepeng@nwpu.edu.cn Abstract Framing is a communication strategy to bias discussion by selecting and emphasizing. Frame detection aims to automatically analyze framing strategy. Previous works on frame de- tection mainly focus on a single scenario or issue, ignoring the special characteristics of frame detection that new events emerge con- tinuously and policy agenda changes dynam- ically. To better deal with various context and frame typologies across different issues, we propose a two-stage adaptation framework. In the framing domain adaptation from pre- training stage, we design two tasks based on pivots and prompts to learn a transferable en- coder, verbalizer, and prompts. In the down- stream scenario generalization stage, the trans- ferable components are applied to new issues and label sets. Experiment results demonstrate the effectiveness of our framework in different scenarios. Also, it shows superiority both in full-resource and low-resource conditions. 1 Introduction Framing is a communication strategy, used to bias the discussion toward a specific stance, by select- ing particular aspects of reality and making them more salient (Entman, 1993; Liu et al., 2019a). It is widely adopted by politicians, the media, and the voting public to seek support, express opin- ions, and advance political agendas (Levendusky, 2013), thus having important implications for pub- lic opinion understanding and policy decision- making (Mendelsohn et al., 2021). Boydstun et al. propose 15 generic frame dimensions based on pol- icy agenda, including economic, morality, and so on, paving the way for frame analysis. In order to analyze the framing strategy automatically, re- searchers explore the task of frame detection. Given a piece of statement related to a topic, frame detection aims to recognize which dimen- sions of frames are employed. Formally, it’s a ∗Corresponding author. Issue Source # label Example Frame abortion tweets 14 Women have the freedom to make their own choices about their reproductive health care. #AbortionLaw protects the safety of women. gun news 9 Florida shooter a troubled loner with white supremacist ties. gun debates 5 The 2nd amend does not say people need guns for self defense ! Fairness & Equality Health & Safety Mental Health Race/Ethnicity Legality, Constitutionality and Jurisprudence Figure 1: A brief landscape of frame detection. These three samples are taken from twitter (Johnson et al., 2017), gvfc (Liu et al., 2019a), fora (Hartmann et al., 2019) respectively. multi-label classification task with a pre-defined label set. Existing researches for automatic frame detection explore different supervised methods including feature-based machine learning, deep neural networks, and fine-tuning pre-trained mod- els (Card et al., 2016; Naderi and Hirst, 2017; Liu et al., 2019b). Despite the success and contribution made …

Content truncated to 500 words