Cross-Directional Pairing Failure in Dependency Relation Classification

Research shows that while splitting BiLSTM representations into forward and backward components improves dependency classification, pairing a forward state with a backward state of a different token (cross-direction) consistently underperforms same-direction pairing. This performance gap increases with token distance and persists even when training only fresh heads on a frozen trunk.