[4Xin2] Poster session 2

[4Xin2-90] Generation of Descriptive Text based on Comparison of the Behavior of Two Time-series Data

〇Yukako Nakano¹, Ichiro Kobayashi¹ (1.Ochanomizu University)

Keywords:Time-series data, Natural Language Generations, Data-to-Text

Data-to-Text is a research of describing numerical data expressed in a variety of formats, such as tables and time-series charts in natural language, and helps people understand the data. Studies describing the numerical analysis results mostly deal with one time-series data. Describing not a single but two time-series data enables us to explain more information like comparison results. This study proposes method that can explain the results of behavioral comparisons of two time-series data. We created three datasets and trained our proposed model that extends the transformer. With the created datasets, we compared the results of training by taking steps and do not. The results showed that taking steps in the training increases the accuracy of generated explanations.