Page 14 - APEC CLIMATE CENTER 2025 Annual Report

APEC CLIMATE CENTER 2025 ANNUAL REPORT

Highlighted tation and enabled the automatic generation of concise, objective natural-language de-

Highlighted Achievements in 2025 Achievements scriptions for probabilistic forecasts, sea surface temperature anomalies, and ENSO-relat-
ed information, all within a consistent narrative and logical structure.
in 2025 Finally, the entire workflow—from data preprocessing and signal detection to lan-

2)
guage-model-based narrative generation and LaTeX-based document assembly—was
integrated into a single automated pipeline. An automatic compilation system was de-
veloped to faithfully reproduce the layout and visual structure of existing Pacific Climate
Highlighted 1. A New Way of Producing Climate Outlooks: A Data Outlook documents. Through the execution of a single script, the full sequence from
data ingestion to final PDF output is completed in order, resulting in a clearly defined
Achievements and Rule-Based Systematization and fully reproducible workflow. This integration established a coherent production

structure in which analysis, narrative generation, and document preparation are seam-
in 2025 ㉖ - Dr. Jin Ho Yoo (jhyoo@apcc21.org) lessly connected.

Climate outlooks serve as core decision-support tools for anticipating future weather-re-
lated risks through the analysis of climate variability on seasonal timescales. In regions
with high climate vulnerability, such as the Pacific Island Countries, climate outlooks are
essential public services that support national climate adaptation strategies and informed
decision-making at the community level. Beyond simply presenting forecast results, cli-
mate outlooks represent high-value information products that synthesize diverse climate
data through expert interpretation and structured narrative, delivering actionable insights
to policymakers and practitioners.

However, traditional climate outlook production processes have relied heavily on the ex-
perience and qualitative judgment of individual experts. As a result, maintaining consis-
tency and reproducibility across the recurring monthly cycle of large-scale data analysis
and document preparation has posed persistent challenges. Variations in interpretation Fig 8 Forecast signals mapped over the subregion masks. Different colors correspond to different
signals (category/strength).
and narrative style depend on the author, while fragmented analysis, writing, and edit-
ing hinder the coherent management of the production process. To address these struc-
tural limitations and strengthen the sustainability and reliability of climate information
services, this initiative transformed the Pacific climate outlook production process into a
data- and rule-based automated framework.

The first step involved establishing a quantitative signal detection and regional match-
ing system that reflects the climatological and geographical characteristics of the Pacific
region. By systematically reviewing climate outlook documents produced over the past
five years, commonly used regional classifications and descriptive conventions were
identified and consolidated. Based on this analysis, the surrounding areas of the Pacific
Island Countries were standardized into 17 core subregions. Numerical signals derived
from probabilistic forecast fields were subsequently mapped to the standardized termi-
nology traditionally used in climate outlook reports—such as ‘Strongly Enhanced,’‘En-
hanced,’ and ‘Slight Tendency’—through stringent quantitative interpretation rules. A pri-
ority-based decision logic combining probability thresholds and spatial coverage ratios
was applied, ensuring consistent signal selection without subjective intervention from
individual forecasters.

1)
In parallel, an offline, large language-model-based narrative automation system was
implemented to operate stably within internal computing environments where external
network access is restricted. A compact language model was adopted to reflect the char-
1) LLM (Large Language Model) 2) LaTeX : Fig 9 Process flow of integrated outlook generation pipeline
acteristics of Pacific Climate Outlooks and their established writing stlye. Additionally, a
An AI language model trained on large A document typesetting system de-
structured “Prompt Header” approach was introduced to feed signal detection results
amounts of text to generate, summa- signed for precise formatting of text
into the model in a controlled manner. This design minimized arbitrary model interpre-
rize, and translate text and mathematical expressions

14 15 15
14