Byte-sized reviewers: exploring review generation through LNRG-S

The development of network technology has led to an explosive growth of user-generated comments. These contents are based on users' experiences after using services and products, reflecting users' evaluations. Automatically generating high-quality and controllable sentiment comments can help creators improve work quality, help service providers enhance service levels, and improve consumers' consumption experience. However, current methods have some limitations. Seq2seq models have difficulty maintaining semantic consistency and generating long texts, LSTM can maintain a memory state and generate coherent text, but is prone to overfitting, generating repetitive text, and having difficulty establishing local coherence. N-gram models text locally but cannot understand the global context. This paper proposes a hybrid architecture, called LNRG-S (LSTM N-gram Reviews Generator with Sentiments), which combines LSTM and N-gram to model both global and local dependencies, and trains separate LSTM generators for positive and negative samples to control sentiment polarity of generated comments. We computed the perplexity of comments generated by different models. LNRG-S has 43.93% and 18.89% lower perplexity than LSTM and n-gram respectively, demonstrating that LNRG-S generates more fluent and coherent text.