KG-BART: Knowledge Graph-Augmented BART for Generative Commonsense Reasoning

Generative commonsense reasoning which aims to empower machines generate sentences with the capacity of over a set concepts is critical bottleneck for text generation. Even state-of-the-art pre-trained language generation models struggle at this task and often produce implausible anomalous sentences. One reason that they rarely consider incorporating knowledge graph can provide rich relational information among concepts. To promote ability generation, we propose novel augmented model KG-BART, encompasses complex relations through produces more logical natural as output. Moreover, KG-BART leverage attention aggregate concept semantics enhances generalization on unseen sets. Experiments benchmark CommonGen dataset verify effectiveness our proposed approach by comparing several strong models, particularly outperforms BART 5.80, 4.60, in terms BLEU-3, 4. also show generated context work background scenarios benefit downstream QA tasks.