MTMS: A Fact-Corrected Summarization model Based on Multitask Learning and Multimodal Fusion

Data-driven large-scale neural network models have now become the dominant paradigm for text summarization tasks. However, the factual inconsistency problem remains a very challenging challenge in the field of text summarization. To alleviate this problem, we introduce multitask learning with multimodal fusion into the text summarization domain and propose the MTMS model. The model effectively models multimodal data fusion by introducing image modal data to assist in correcting factual errors, averaging multi-directional noise through multi-task learning, and by a special hierarchical attention fusion mechanism. Experiments show that the MTMS model is significantly effective in correcting factual errors without sacrificing ROUGE scores.