Development of a CGAN-Based Method for Aspect Level Text Generation: Encouragement and Punishment Factors in the Aspect Knowledge

Text mining systems may benefit from the use of automated text generation, especially when dealing with limited datasets and linguistic resources. Most successful text generation approaches are generic rather than aspect-specific, resulting in relatively inaccurate and similar sentences in different aspects. The present study proposes a solution to this problem by extracting aspect knowledge from relevant topics and creating the correct phrase based on the Conditional Generative Adversarial Network (CGAN) for each aspect. The proposed method produces sentences using an auxiliary dataset that cannot be distinguished from genuine sentences by the discriminator. In order to generate an auxiliary dataset, aspect-based information from datasets related to the target concept is extracted. To further improve the accuracy, the generator is encouraged or punished depending on the similarity with the training corpus. Two datasets in English and Persian are used to evaluate the performance of the proposed text generation method. The results show that adding similar aspects to the auxiliary dataset improves the quality of the generated sentences. In addition, encouragement leads to more accurate sentences, while punishment leads to more varied sentences.

Keywords:

Deep Learning , Text Generation , Conditional Generative Adversarial Network , Aspect , Long Short-Term Memory