Synthetic Text Generation for Sentiment Analysis

• Umar Maqsud ^[1]
  1. [1] Technical University of Berlin

     Technical University of Berlin

     Berlin, Stadt, Alemania

     
• Localización: 6th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis WASSA 2015: Workshop Proceedings : 17 September 2015 Lisboa, Portugal / Alexandra Balahur Dobrescu (ed. lit.), Erik van der Goot (ed. lit.), Piek Vossen (ed. lit.), Andrés Montoyo Guijarro (ed. lit.), 2015, ISBN 978-1-941643-32-7, págs. 156-161
• Idioma: inglés
• Enlaces
  • Texto completo (pdf)
• Resumen

  • Natural language is a common type of input for data processing systems. Therefore, it is often required to have a large testing data set of this type. In this context, the task to automatically generate natural language texts, which maintain the properties of real texts is desirable. However, current synthetic data generators do not capture natural language text data sufficiently. In this paper, we present a preliminary study on different generative models for text generation, which maintain specific properties of natural language text, i.e., the sentiment of a review text. In a series of experiments using different data sets and sentiment analysis methods, we show that generative models can generate texts with a specific sentiment and that hidden Markov model based text generation achieves less accuracy than Markov chain based text generation, but can generate a higher number of distinct texts.