Multiple POS Dependency-Aware Mixture of Experts for Frame Identification

Frame identification, which is finding the exact evoked frame for a target word in given sentence, fundamental and crucial prerequisite semantic parsing. It generally seen as classification task words, whose contextual representations are usually obtained using neural network like BERT an encoder, enriched with joint learning model or knowledge of FrameNet. However, distinction at fine-grained level, such delicate differences information syntax PropBank roles caused by different parts-of-speech (POS) neglected. We propose Multiple POS Dependency-aware Mixture Experts(MPDaMoE) that integrates five types information, consisting syntactic words nominal, adjectival, adverbial, prepositional, role only verbal. To better learn Experts employed, every expert Graph Convolutional Network, to incorporate dependency words. Our outperforms state-of-the-art models experiments on two benchmark datasets, shows its effectiveness.