0xbt

On this paper, we propose a novel Cluster-to-Cluster Generation framework for Data Augmentation of slot filling, named C2C-GenDA. On this paper, we study the info augmentation for slot filling activity that maps utterances into semantic frames (slot kind and slot value pairs). In this paper, we set up a novel SLU job, the few-shot noisy SLU, with current public datasets. To additional encourage the range of the generated utterances, we suggest two novel mechanisms: (1) Duplication-conscious Attention that attends to the existing expressions to keep away from duplicated technology for every decoding step. For each semantic frame, we use a Cluster2Cluster (C2C) model to generate new expressions from present utterances. We believe the ensemble nature of ProtoNets advantages the mannequin robustness, and the simplicity of Proto’s model structure is also useful within the few-shot noisy situation. We further propose a ProtoNets primarily based approach, Proto, to construct IC and SL classifiers with few noisy examples.