📄 Cherish at SemEval-2026 Task 2: Enhancing RoBERTa-Based Models for Emotional Valence and Arousal Prediction in Ecological Essays with Personalized PLoRA and Temporal Embeddings

Summary

This SemEval-2026 Task 2 system description addresses predicting self-reported emotional valence and arousal from longitudinal ecological essays and short feeling words. The core difficulty is that the same wording can carry different affective intensity depending on the writer and on how much time has passed between entries. Team Cherish starts from RoBERTa-large and keeps its general semantics largely intact. On top of it they add three components: Personalized LoRA (PLoRA), which injects user-conditioned low-rank updates; a learned embedding of the time gap between a user's consecutive texts; and separate regression heads for valence and arousal. Training uses the organizers' EMA dataset with user-grouped cross-validation. In Subtask 1 the system ranked 13th of 29 teams. For Subtask 2a (forecasting change), the author added a GRU autoregressive decoder, but it produced constant outputs, so the official correlation was undefined. The paper is useful mainly as a template for parameter-efficient personalization combined with explicit time features. It also offers a candid look at how such personalized forecasters can collapse.

Key points

• Architecture:
  • Backbone is RoBERTa-large with attention-mask mean pooling (1024-dim output).
  • LoRA adapters (r=8, α=32) sit on the query and value projections in all 24 layers.
  • A user vector is added through a separate projection.
  • Personalized Dropout is used during training, and the personalization signal is zeroed for unseen users.
  • A Δt MLP (1→64) output is concatenated with the text vector and projected from 1088 to 128 dimensions.
• Data:
  • The full dataset contains 5,285 EMA observations from 182 users, collected 2021–2024.
  • Training uses 2,764 observations from 137 users.
  • Preprocessing removed keysmash noise and duplicate entries.
  • Validation used 5-fold GroupKFold grouped by user ID.
• Subtask 1 results:
  • Composite Pearson r was 0.596 for valence and 0.505 for arousal.
  • Between-user r was 0.648 (valence) and 0.616 (arousal).
  • Within-user r was 0.538 (valence) and 0.375 (arousal).
  • Continuous predictions were mapped to discrete values by thresholds before submission.
• Subtask 2a results:
  • Official correlations were NaN because predictions had zero variance.
  • MAE was 1.565 for valence and 2.130 for arousal.
• Diagnosed causes of the 2a collapse:
  • Heavy teacher forcing in the GRU decoder; scheduled sampling is proposed as the fix.
  • No safeguard for user IDs outside the embedding table; PDropout and an "Unknown" embedding slot are proposed.
  • The paper does not clarify why PDropout is described as part of the method yet recommended as future work.