How LSTM Encodes Syntax: Exploring Context Vectors and Semi-Quantization on Natural Text

Proceedings of the 28th International Conference on Computational Linguistics, pages 396–406 Barcelona, Spain (Online), December 8-13, 2020 396 Predicting Stance Change Using Modular Architectures Aldo Porco Purdue University aporco@purdue.edu Dan Goldwasser Purdue University dgoldwas@purdue.edu Abstract The ability to change a person’s mind on a given issue depends both on the arguments they are presented with and on their underlying perspectives and biases on that issue. Predicting stance changes requires characterizing both aspects and the interaction between them, especially in realistic settings in which stance changes are very rare. In this paper, we suggest a modular learning approach, which decomposes the task into multiple modules, focusing on different aspects of the interaction between users, their beliefs, and the arguments they are exposed to. Our experiments show that our modular approach archives sig- nificantly better results compared to the end-to-end approach using BERT over the same inputs. 1 Introduction One of the main drivers of social interaction is convincing people to adopt new ideas and perspectives. Understanding how to make compelling arguments that would achieve this goal has been studied exten- sively in psychology and the social sciences (Fogg and B.J., 2003; Popkin, 1991), and more recently by the NLP community. Most of these works focus on analyzing argumentative text in isolation (Habernal and Gurevych, 2016a; Potash and Rumshisky, 2017; Persing and Ng, 2017; Gleize et al., 2019). While the quality of arguments can potentially be judged in isolation based on their merits, their effectiveness, when it comes to convincing readers, has to be considered in conjunction with the subjective perspectives and biases of these users. For example, social psychology studies have shown that ideological stances are highly correlated with different moral arguments preferences (Haidt and Graham, 2007; Graham et al., 2009; Graham et al., 2012; Johnson and Goldwasser, 2018), while Lukin et al., (2017) studied the re- lation between personality type and factual vs. emotional arguments. Durmus and Cardie (2018) studied linguistic properties of convincing arguments, conditioned on users’ political and religious convictions. This work aims to study the effectiveness of debate arguments, with respect to the biases of users read- ing them. Our goal is to mimic a realistic setting, in which only partial information about these users and their behaviors is available. To accommodate this setting, we use data from the website debate.org, used in several previous studies (Durmus and Cardie, 2018; Durmus and Cardie, 2019; Luu et al., 2019; Pacheco and Goldwasser, 2020), and formulate a new classification task, ChangeMyStance, (CMS)1, pre- dicting whether a specific user would change their stance after being exposed to two conflicting texts: one that does not support the voting user’s views and another one that does. As people do not easily change their minds when exposed to other perspectives, this formulation results in a highly challenging task with only about 6% of users persuaded by the arguments. This formulation is different than the one proposed by Durmus and Cardie (2018), which also studied users’ bias, as explained in Section 4. Our main …

Content truncated to 500 words