Postprocessing System for Word-Filling and Quantifiers Based on Chinese Story Generation

Abstract

In recent years, many fields have been related to introducing artificial intelligence to natural language generation. Although these natural language models have excellent results and generate smooth sentences, they are still not effective learning features such as character relationships, especially in the Chinese language. When a sentence is generated, it is necessary to pay attention to the following words to correctly predict and generate, such as quantifiers, which causes the generated words to be inappropriate and then affects the generation of subsequent words. Therefore, we developed a set of attention mechanism enhancement models, aiming at the generative language model that controls the ordering of generated speech parts, revising the different mechanisms of character fighting relationship and quantifier design, and adopting the traditional classification model one-against-all support vector machine training. The results show that the generated sentences are arranged in the same order as the original ones so the generation can be reliably controlled.