RNZ Climate Corpus Exploratory Analysis

An exploratory corpus-linguistics project examining how climate change and methane are represented in RNZ national news reporting.

The analysis compares the RNZ Climate National Corpus with a much larger general RNZ reference corpus. It asks two linked questions:

1. Which words are unusually prominent in national climate reporting relative to general RNZ news?
2. Which words co-occur most strongly with methane, and what discourse contexts do those patterns indicate?

The project combines quantitative corpus statistics with cautious discourse interpretation. It focuses on repeated language patterns rather than inferring author intentions.

The target corpus contains:

• 6,104 documents;
• 3,962,664 tokens;
• RNZ national climate-related articles published from 2008 to 2024.

The analysis was completed with the open-source Conc Python library developed for reproducible corpus analysis in Jupyter notebooks.

The reported statistics, interpretations, and selected figures in this repository come from Sahil Mann's completed Corpus Exploratory Analysis and its recorded outputs. The previous-semester university corpus environment is no longer available, so this portfolio package does not claim a new execution.

The companion notebook reconstructs the documented Conc workflow against the exact corpus names and provides aggregate-only export and reconciliation steps for an authorised future rerun. It does not contain RNZ article text or another student's analysis.

• Frequency records how often a token occurs.
• Normalised frequency supports comparison between differently sized corpora.
• Log Ratio measures the size of the difference between target and reference frequencies.
• Log Likelihood records statistical evidence for the observed difference.
• LogDice measures collocation strength around methane.
• Methane collocates use a five-token window on each side and minimum collocate frequency of five.

The notebook labels terms used to construct the climate corpus and produces a sensitivity view that excludes them. This prevents expected collection terms from dominating the substantive interpretation.

Prominent non-collection terms include ets, methane, fossil, coal, 2050, targets, farmers, agriculture, and renewable. Together, they connect RNZ climate reporting with:

• policy instruments and emissions-reduction strategies;
• long-term targets and future dates;
• agriculture, livestock, and biogenic methane;
• fossil energy and renewable alternatives.

Strong methane collocates include oxide, biogenic, nitrous, livestock, reduction, emissions, targets, animals, reduce, 2030, cows, and cattle. The pattern places methane at the intersection of scientific terminology, farming, livestock emissions, and policy targets.

.
|-- README.md
|-- requirements.txt
|-- assets/
|   |-- rnz_climate_keywords_log_ratio.png
|   `-- rnz_methane_collocates_logdice.png
|-- notebooks/
|   `-- rnz_climate_corpus_analysis.ipynb
`-- validation/
    |-- reported_keyword_results.csv
    `-- reported_methane_collocates.csv

The full rerun requires access to these Conc corpus directories:

rnz-climate-national.corpus
rnz.listcorpus

They are available in the University of Canterbury DIGI405 JupyterHub environment under /srv/corpora/. The notebook accepts another authorised location through RNZ_CORPUS_ROOT.

python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
jupyter lab notebooks/rnz_climate_corpus_analysis.ipynb

On an authorised execution, the notebook exports full aggregate keyword and collocate tables and reconciles selected rows against the completed report. This makes package-version or corpus-version differences visible without replacing the report-backed results presented here.

RNZ article text and the server-hosted corpora are not included. The repository contains analysis code, aggregate statistics, and charts only. Concordance lines are intentionally not exported because they reproduce article text.

• Keywords show relative prominence, not importance in an absolute sense.
• Collocation identifies repeated co-occurrence, not causation or institutional intent.
• Corpus collection terms can inflate expected words, so substantive interpretation uses a sensitivity view.
• The analysis describes RNZ reporting in this corpus and time period; it should not be generalised to all New Zealand media.
• A more complete study would test change over time, article sections, authorship, and source diversity.

Python, Jupyter, corpus data structures, comparative keyword analysis, Log Ratio, Log Likelihood, LogDice, collocation analysis, discourse interpretation, sensitivity analysis, aggregate-result export, validation, and responsible handling of copyrighted text.