Prior work showed that automatically induced emotion labels can help predict the temporal orientation of social media posts, but it was limited to a small set of discrete emotions (e.g., joy, anger, sadness, fear). In practice, affect is often better captured along continuous dimensions such as valence and arousal, or through finer-grained emotion inventories. In this paper we investigate whether richer affect signals can improve tweet-level and user-level prediction of temporal orientation (past, present, future). We construct a weakly supervised pipeline that (i) assigns temporal-orientation pseudo-labels using multiple textual heuristics and label aggregation, and (ii) derives dimensional affect features by projecting posts into valence–arousal–dominance (VAD) space and by distantly supervising fine-grained emotion labels. We then train multi-task neural classifiers in which temporal orientation is the primary task and affect dimensions are auxiliary tasks. Experiments on an English Twitter corpus of UK users show that (1) dimensional affect is more consistently helpful in cases where temporal cues are implicit or underspecified, (2) user-level aggregation further amplifies the gains, and (3) combining discrete and dimensional affect gives the best macro-F1. Our analysis suggests that temporal orientation in social media is partly mediated by affective stance, and that future work should account for socio-cultural shifts and platform variation.
Prior work showed that automatically induced emotion labels can help predict the temporal orientation of social media posts, but it was limited to a small set of discrete emotions (e.g., joy, anger, sadness, fear). In practice, affect is often better captured along continuous dimensions such as valence and arousal, or through finer-grained emotion inventories. In this paper we investigate whether richer affect signals can improve tweet-level and user-level prediction of temporal orientation (past, present, future). We construct a weakly supervised pipeline that (i) assigns temporal-orientation pseudo-labels using multiple textual heuristics and label aggregation, and (ii) derives dimensional affect features by projecting posts into valence–arousal–dominance (VAD) space and by distantly supervising fine-grained emotion labels. We then train multi-task neural classifiers in which temporal orientation is the primary task and affect dimensions are auxiliary tasks. Experiments on an English Twitter corpus of UK users show that (1) dimensional affect is more consistently helpful in cases where temporal cues are implicit or underspecified, (2) user-level aggregation further amplifies the gains, and (3) combining discrete and dimensional affect gives the best macro-F1. Our analysis suggests that temporal orientation in social media is partly mediated by affective stance, and that future work should account for socio-cultural shifts and platform variation.
