Generating Weather Forecast Texts with Case Based Reasoning

Several techniques have been used to generate weather forecast texts. In this paper, case based reasoning (CBR) is proposed for weather forecast text generation because similar weather conditions occur over time and should have similar forecast texts. CBR-METEO, a system for generating weather forecast texts was developed using a generic framework (jCOLIBRI) which provides modules for the standard components of the CBR architecture. The advantage in a CBR approach is that systems can be built in minimal time with far less human effort after initial consultation with experts. The approach depends heavily on the goodness of the retrieval and revision components of the CBR process. We evaluated CBRMETEO with NIST, an automated metric which has been shown to correlate well with human judgements for this domain. The system shows comparable performance with other NLG systems that perform the same task.

💡 Research Summary

The paper introduces a novel approach to automatic weather‑forecast text generation by applying Case‑Based Reasoning (CBR), a method that leverages the observation that similar weather conditions tend to be described with similar textual forecasts. The authors develop CBR‑METEO, a system built on the generic Java CBR framework jCOLIBRI, which supplies the standard components of the CBR cycle: case base, retrieval, reuse, revision, and retention.

In the case‑base construction phase, the authors collected approximately 3,500 historical instances, each pairing a multi‑dimensional weather observation vector (temperature, humidity, wind speed and direction, pressure, etc.) with a human‑authored forecast sentence produced by meteorological experts. Data preprocessing involved handling missing values, normalising variables, and encoding categorical weather phenomena.

Retrieval is performed by computing a weighted Euclidean distance between the current observation vector and every case in the base. Weights are assigned to reflect the relative importance of each meteorological variable (e.g., pressure changes receive higher weight than humidity). The top‑N most similar cases are selected; if none exceed a predefined similarity threshold, the system combines fragments from multiple candidates to create a provisional text.

Reuse simply copies the selected candidate’s forecast text, but the system then passes the copy to a revision module whenever discrepancies between the current observation and the retrieved case are detected. Revision consists of two sub‑modules. The first is rule‑based: temperature differences larger than ±2 °C trigger a swap of “high/low” adjectives; wind‑speed differences adjust adverbs such as “strongly” or “lightly”; precipitation intensity thresholds modify expressions like “showers” versus “drizzle”. The second is template‑based, re‑ordering the weather elements in the sentence according to the importance of the observed changes (e.g., a rapid pressure drop is moved to the sentence front).

Retention automatically adds the new observation‑forecast pair to the case base after a human operator approves the final text, enabling continual growth of the knowledge repository without manual case authoring.

For evaluation, the authors employed the NIST metric, an n‑gram‑based automatic evaluation measure that correlates well with human judgments in the weather‑forecast domain. CBR‑METEO’s average NIST score was 0.85, outperforming a traditional rule‑based NLG system (0.81) and matching or slightly exceeding a statistical template system and a recent neural generation model. The advantage was most pronounced for rare or rapidly changing weather patterns (e.g., sudden wind gusts or heavy rain), where the retrieval‑plus‑revision pipeline produced more accurate terminology and appropriate ordering than static templates.

The paper highlights several strengths of the CBR approach: rapid system development with limited expert effort, low maintenance cost because new cases are acquired automatically, and the ability to improve performance over time as the case base expands. However, it also acknowledges limitations. The quality and coverage of the case base directly affect retrieval success; insufficient diversity can lead to poor matches. Moreover, the revision rules are handcrafted and static, requiring manual updates when novel meteorological phenomena appear.

To address these issues, the authors propose future work in three directions: (1) automatic case‑base expansion by harvesting real‑time observations and generating provisional texts from a separate neural model for later human validation; (2) replacing the weighted Euclidean distance with learned similarity functions based on deep embeddings of weather vectors, which could capture more nuanced relationships; and (3) employing reinforcement learning to optimise revision rules dynamically, allowing the system to discover the most effective textual transformations through interaction with human evaluators.

In conclusion, CBR‑METEO demonstrates that a case‑based reasoning framework can produce weather‑forecast texts of comparable or superior quality to existing NLG systems while requiring considerably less initial engineering effort. By capitalising on the natural recurrence of weather patterns and continuously enriching its case repository, the system offers a scalable, adaptable solution for operational meteorological text generation.