Stance Prediction from Multimodal Social Media Data

Proceedings of Recent Advances in Natural Language Processing, pages 242–248 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_027 242 Stance Prediction from Multimodal Social Media Data La´ıs C. L. Cavalheiro Matheus C. Pavan School of Arts, Sciences and Humanities University of S˜ao Paulo Av Arlindo Bettio 1000, S˜ao Paulo, Brazil {laiscarraro,matheus.pavan,ivandre}@usp.br Ivandr´e Paraboni Abstract Stance prediction – the computational task of inferring attitudes towards a given target topic of interest – relies heavily on text data provided by social media or similar sources, but it may also benefit from non-text information such as demographics (e.g., users’ gender, age, etc.), network structure (e.g., friends, followers, etc.), interactions (e.g., mentions, replies, etc.) and other non-text properties (e.g., time informa- tion, etc.). However, so-called hybrid (or in some cases multimodal) approaches to stance prediction have only been developed for a small set of target languages, and often making use of count-based text models (e.g., bag-of-words) and time-honoured classification methods (e.g., support vector machines). As a means to fur- ther research in the field, in this work we intro- duce a number of text- and non-text models for stance prediction in the Portuguese language, which make use of more recent methods based on BERT and an ensemble architecture, and ask whether a BERT stance classifier may be en- hanced with different kinds of network-related information. 1 Introduction Standard stance prediction concerns the inference of for/against attitudes towards a target topic of interest from text data. The task may be seen as analogous to sentiment (e.g., positive or negative) analysis, but stance and sentiment do not neces- sarily correlate (Aldayel and Magdy, 2021). For instance, consider the following statement: ‘People who refuse the vaccine should be banned from entering the premises’ In this example, given the intended target ‘vacci- nation’, the statement suggests a favourable stance. Still, the overall sentiment (particularly in the use of the word ‘banned’) may be regarded as being more on the negative side. Computational models of stance prediction, which often take social media text as an input, have been applied to a wide range of topics, in- cluding moral or social issues (Pavan et al., 2023; Geiss et al., 2022), politics (Darwish et al., 2017; Lehmann and Derczynski, 2019; Cignarella et al., 2020), and others. The task has become particu- larly popular in the field since the SemEval-2016 shared task (Mohammad et al., 2016) and accompa- nying corpus, focusing on stance prediction from Twitter text in the English language. In addition to using text data (Zhang et al., 2020; Allaway et al., 2021; Pavan et al., 2020), recent work in stance prediction has addressed the use of non-text data as well (Aldayel and Magdy, 2021). Studies of this kind are largely motivated by the notion of homophily (McPherson et al., 2001), that is, the concept of ‘similarity breeds connection’, and take into account a range of well-known non- linguistic stance predictors. These include, for in- stance, the use of demographics information (e.g., users’ gender, age, etc.) (Lehmann and Derczyn- ski, 2019; Geiss et …

Content truncated to 500 words