[3Rin4] Interactive 1

[3Rin4-07] Relieving the Reference Time Un-Alignment in Generating Trend Report for Time Series Data

〇Yumi Hamazono^1,4, Yui Uehara⁴, Hiroshi Noji⁴, Yusuke Miyao^2,4, Hiroya Takamura^3,4, Ichiro Kobayashi^1,4 (1.Ochanomizu University, 2.The University of Tokyo, 3.Tokyo Institute of Technology, 4.National Institute of Advanced Industrial Science and Technology)

Keywords:Neural language processing, Time-series data, Multi-modal

The necessity of the technology that automatically generates text describing data so that non-experts can interpret large and complex data is increasing. In recent study, the technology has shown high accuracy by utilise end-to-end learning. When creating the dataset, the alignment of the data and text is obtained by manual annotation or other data that accompanies them, but the mismatch of the references point of the data and text possibly occurs.
In this study, we propose a method to relieve the inconsistency of the reference points, using an example of resolving reference time un-alignment in trend report generation for time-series data . Specifically, we worked on the relieving using not only the latest data of the time when the report was announced, but also the data in several time step before.