Popularity Prediction of Online Petitions using a Multimodal DeepRegression Model

Popularity Prediction of Online Petitions using a Multimodal Deep Regression Model Kotaro Kitayama♠ Shivashankar Subramanian♥ Timothy Baldwin♥ ♠ Tohoku University ♥ The University of Melbourne kitayama@ecei.tohoku.ac.jp {shivashankar.subramanian,tbaldwin}@unimelb.edu.au Abstract Online petitions offer a mechanism for peo- ple to initiate a request for change and gather support from others to demonstrate support for the cause. In this work, we model the task of petition popularity using both text and image representations across four different languages. We evaluate our proposed approach using a dataset of 75k petitions from Avaaz.org, and find strong complementarity between text and images. 1 Introduction A petition is a formal request for change or an action to any authority, co-signed by a group of supporters (Ergazakis et al., 2012). The targets of petitions are usually government agencies and business organizations. In this work we study petitions from Avaaz.org, a popular petition platform available across six con- tinents, with support for seventeen different lan- guages. Avaaz provides a platform for petitions and funding campaigns. An example petition is given in Figure 1, opposing the Anti-Counterfeiting Trade Agreement (ACTA) and supporting a free and open internet. The popularity of a petition, in terms of the number of signatures it attracts, is critical to its success, and predicting popularity can help petition organizers to enhance their engagement strategy by optimising the petition content. In particular in this work, we target the task of predicting petition popularity at the time of submission (independent of any social media or direct signature signal). While existing work on petitions has focused on their text content (Elnoshokaty et al., 2016; Subra- manian et al., 2018), images are also a key ingre- dient.1 Additionally, despite petitions being pop- ular in many different languages, there has been 1https://secure.avaaz.org/en/ community_petitions/how_to_create_ petition/ no work on multilingual modeling. From a so- cial science viewpoint, multilingual analysis can contribute to an understanding of issues present in different languages (or regions). Previous research has shown that, other than peti- tion content, metadata is also effective in modelling its popularity. Elnoshokaty et al. (2016) showed that the category of a petition has an influence on its popularity and success, e.g., human trafficking related petitions get more signatures than health re- lated ones. Vidgen and Yasseri (2019) investigated the interaction between topics and geographic fea- tures, and showed that some issues receive broad geographic support (e.g., law & order, work & play) but others are far more local (e.g., animals & nature, driving). Since we aim to model petition popularity across multiple countries, we study the utility of the author’s country information. Our contributions in this work are as follows: (1) we propose a multimodal regression approach for petition popularity prediction task using text and image features; (2) we experiment in both mono- and multi-lingual data settings, to evaluate the im- pact of training data from other languages; and (3) we develop a novel multimodal, multilingual dataset for the task.2 2 Related Work The majority of work on modeling petition popu- larity has focused on predicting …

Content truncated to 500 words