Generating Weather Comments from Meteorological Simulations

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 1462–1473 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 1462 Generating Weather Comments from Meteorological Simulations Soichiro Murakami† Sora Tanaka† Masatsugu Hangyo§ Hidetaka Kamigaito† Kotaro Funakoshi† Hiroya Takamura†,∗ Manabu Okumura† †Tokyo Institute of Technology §Weathernews Inc. ∗Artificial Intelligence Research Center, AIST {murakami,tanaka,kamigaito,funakoshi}@lr.pi.titech.ac.jp, hangyo@wni.com, {takamura,okumura}@pi.titech.ac.jp Abstract The task of generating weather-forecast com- ments from meteorological simulations has the following requirements: (i) the changes in nu- merical values for various physical quantities need to be considered, (ii) the weather com- ments should be dependent on delivery time and area information, and (iii) the comments should provide useful information for users. To meet these requirements, we propose a data- to-text model that incorporates three types of encoders for numerical forecast maps, obser- vation data, and meta-data. We also introduce weather labels representing weather informa- tion, such as sunny and rain, for our model to explicitly describe useful information. We conducted automatic and human evaluations. The results indicate that our model performed best against baselines in terms of informative- ness. We make our code and data publicly available1. 1 Introduction Numerical weather prediction (NWP), a method for weather forecasting that uses mathematical models of the atmosphere, oceans, and observations, has become a mainstream tool for supporting today’s weather forecasts around the world. Weather fore- casters obtain numerical outputs from the simula- tion models and use their scientific knowledge and historical data to come up with forecast comments such as “sunny and sometimes cloudy”. However, writing local or personalized weather comments for end users is labor intensive and requires a solid knowledge of meteorology. Therefore, the task of generating weather-forecast comments has tradi- tionally been addressed in the field of data-to-text generation (Goldberg et al., 1994; Belz, 2007). In this paper, we focus on the task of generating weather-forecast comments from meteorological 1https://github.com/titech-nlp/ pinpoint-weather Cloud cover Air pressure Air temperature Precipitation Delivery time: 05:51 a.m. on 06 April, Tokyo Today patches of blue sky will appear, but the sky will become cloudy and it will gradually start to rain in the evening. Please bring an umbrella wh- en you go out, even if it’s not raining. Figure 1: A weather comment written by a meteorolog- ical expert and simulation results from NWP models simulations. While previous studies have mainly focused on database records and tables (Sripada et al., 2004; Liang et al., 2009), which are modified results by experts based on their local knowledge (Reiter et al., 2005), we use raw simulation results of NWP models as inputs for text generation. This is closer to the real-world scenarios, in which mete- orological specialists describe weather comments by interpreting such numerical data. We believe it will be more helpful for less experienced forecast- ers. There has been little research on generating descriptions from a sequence of raw numerical data even in the data-to-text generation (Gatt and Krah- mer, 2018). We illustrate the three characteristic problems of weather-comment generation …

Content truncated to 500 words