Research And Implementation Of Poetry Generation Robot Based On Deep Learning

Ancient Chinese poetry is a treasure of Chinese culture.In recent years,the task of Chinese jueju generation has gradually become an important research topic in the field of NLP.Generally,there are still four problems in the field of Chinese poetry generation:the lack of attempt of using Transformer based model to solve the problem of generation,the poor controllability of poetry generation,the lack of research on acrostic poetry generation,and the lack of research on compressing and deploying deep learning algorithms to realize a offline chatbot.In order to solve those three algorithm problems and one engineering problem,the following work has been done:First,this paper propose and implement TransPoetc algorithm for the task of generating ordinary ancient Chinese poetry.Based on the BART model,a Gaussian distribution is introduced to model the prior distribution,which improves the generation diversity.Compared with the baseline model,experiments and manual evaluation show that TransPoetc is able to generate poems with better overall quality.To the best of our knowledge,this is the first attempt to combine pretrained model with CVAE structure in this field.Secondly,for the task of sentiment controllable Chinese poetry generation,two algorithms are proposed and implemented,sentEnhancedPoet and multiVAEPoet,which effectively solves the problem in two different ways.The results of human evaluation show that the two algorithms are 3.4 and 4.6 percentage points higher than the best baseline model in the whole poem granularity and sentence granularity respectively.To the best of our knowledge,this is the first attempt to use multiple VAE structure to solve controllable generation problem in this field.Thirdly,the headHiderPoet algorithm is proposed and implemented for the task of generating acrostic poems.Besides the typical CVAE structure,the algorithm successfully improve the semantic consistency of the generated poems and the acrostic words by introducing an adversarial network.Ablation study shows that the coherence and ideographic degree of the poems generated by the advanced model are improved evidently.To the best of our knowledge,this is also the first attempt in this field to solve the problem of the acrostic poems by adversarial training.Finally,all the above algorithms are implemented on NVIDIA Jetson Nano.The offline speech recognition is realized by compiling Kaldi offline speech recognition structure and speech synthesis tools.The models are compressed by transferring all PyTorch versions into the TensorRT counterparts,within which the GRU operator is contributed to official torch2trt project.All of the above generation algorithms are fully realized on the embedded development board.The system test results show that the robot realizes the function of poetry generation through multimodal human-computer interaction,and the resource consumption and algorithm reasoning time are controlled in an acceptable time range.