Team BOBW (Best Of Both Worlds) at SemEval-2026 Task 3: Modular Cross-Attention Encoders for Dimensional Aspect-Based Sentiment Analysis

AbstractThis paper presents our system for SemEval-2026 Task 3, which identifies four-part opiniondetails in product reviews. We used a sequenceof pairs of BERT encoder models connectedby cross-attention layers. The cross-attentionmechanism provided marginally better resultsthan a self-attention equivalent, failing to show-case a significant improvement. Error propaga-tion through the pipeline hurt the correctness ofthe outputs, with certain stages collapsing thescores. The pipeline architecture’s performancewas largely independent of model size, sug-gesting that small modular encoders for down-stream tasks are an efficient alternative to largedecoder models. Our best model got a cF1score of 0.53 on restaurant data and 0.26 onlaptop data.