Abstract
This paper describes a natural language generation method to verbalize the trends of observed time-series data. As an example of time-series data, we focus on Nikkei Stock Average and develop the method based on its data. By employing Symbolic Aggregation Approximation (SAX), the dimension reduction and pattern extraction are applied to the observed data. The correspondent relation between the extracted patterns and the semantic labels representing the content of linguistic descriptions about the observed data is learned by means of log linear model, and therefore the patterns of time-series data are identified. For each semantic label, a bi-gram model is built from collected corpus which explains the trends of Nikkei Stock Average. A sentence explaining the stock trends is generated by finding the most likely combination of words from the bi-gram model by means of dynamic programming.
