Updated 1 November 2026

Methodology

How speeches become sixteen indices. Twelve are tagged sentence by sentence by a language model against a fixed rubric and checked against human coding; four are computed from the prose.

The framework

The Index measures speech on sixteen indices in four families of four. Content records what a sentence is about, Actors who is in it and how they are cast, and Style how it is said. These twelve indices are tagged sentence by sentence. Prose holds four scores computed from the text itself. Content and Actors describe the world the speaker builds; Style and Prose describe the speaker's craft. Every index can be computed on any collection of speeches: one leader's career, one country or a corpus such as the UN General Debate.

Unit of analysis

Every speech is split into sentences, and each sentence is tagged on its own. A sentence has clear boundaries and is usually short enough to carry one dominant frame. Quasi-sentences (sub-sentence units split at each distinct argument) are not used: their boundaries are subjective and coders rarely segment a long sentence the same way. Where a sentence carries more than one frame, it is tagged by its dominant one.

Classification

All twelve sentence indices are returned by a single call to a frontier language model per sentence, against a fixed rubric. Each index gets exactly one option or its Null. Speaker, date and other metadata are withheld, so the tag comes from the sentence alone. Tags are stored with the sentence text in each speech's record.

Validation

Before the full corpus is tagged, a pilot of about 5,000 sentences is double-coded by people in two passes of six indices each, not one pass of twelve, so that fatigue does not depress agreement on the later indices. The pilot sets each index's agreement with the model and produces a correlation matrix across indices. Any pair correlating above about 0.7 is cut or merged. The pairs to watch are Arena × Capability (related by design), Emotion × Role and Populism × Role.

Aggregation

A speech's profile on an index is the share of its sentences carrying each option, with Null included so the shares sum to 100 per cent. A year's profile is the mean of speech-level shares across that year's speeches, each speech counting equally, so a long speech does not outweigh a short one. The standard deviation across speeches is reported with every mean: a narrow band means the speeches agree, a wide one that the average hides a split. For ordered indices (Horizon, Conviction), a speech or year can also be summarised as its mean position on the scale.

Content: what the speaker talks about

Arena

The broad sphere of national life a sentence belongs to. Every sentence has an Arena, so it is the denominator the other indices are read against.

Capability

The lever of national power a sentence discusses, in the nine domains of the GINC National Capability Framework, grouped as hard, soft and economic power.

Scope

How far a sentence reaches beyond the speaker's borders.

Horizon

Where in time a sentence points, on one ordered seven-step scale with the present at zero. Distance is given in broad bands rather than years, because most references to the future are vague.

Actors: who is in the story and how they are cast

Actor

Whose actions a sentence is about. Always read with Role.

Role

The part a sentence gives its main character, as if the speech were a story. Null is shown as None.

Stance

Where the speaker stands towards the audience.

Populism

Whether a sentence pits a virtuous people against a corrupt elite, as the ideational school defines populism. Speech-level results can be benchmarked against the Global Populism Database where speakers overlap. Null is shown as None.

Style: how it is said

Expression

The speech act a sentence performs, after Searle's taxonomy.

Conviction

How sure the speaker sounds: the force of the claim, not its content or truth.

Emotion

The feeling a sentence appeals to in the listener, not the speaker's own state. Null is shown as Neutral.

Foundation

The moral value a sentence invokes to claim right or wrong, after Moral Foundations Theory. Provisional: it is dropped if the pilot shows it collapsing into two foundations or tracking Emotion too closely. Null is shown as None.

Prose: the texture of the writing

Prose is computed, not tagged, and gives continuous scores instead of option shares. Each score is computed per sentence. A speech's score is the mean across its sentences, and year and corpus figures weight each speech by its word count. Empty and single-token sentences are skipped.

Readability

Flesch Reading Ease (1948), applied to each sentence: 206.835 − 1.015 × words per sentence − 84.6 × syllables per word. Runs from 0 to 100, and higher is easier. Syllables are counted with a vowel-group heuristic and a silent-e correction.

Grade Level

Flesch–Kincaid Grade Level (1975), from the same inputs: 0.39 × words per sentence + 11.8 × syllables per word − 15.59. This is the years of schooling needed to follow the prose, and lower is simpler. Very short sentences of one-syllable words can score below zero, and these values are kept.

Syntax

Mean dependency distance: how far, in words, each word sits from the word it attaches to, averaged over the non-root tokens of the sentence's spaCy en_core_web_sm parse. Higher means more nested and more complex.

Vocabulary

The mean wordfreq Zipf frequency of content words (nouns, proper nouns, verbs, adjectives and adverbs). The Zipf scale is log₁₀ of occurrences per billion words, on which “the” is about 7. The site charts it as rarity, 7 − Zipf, so higher means rarer.

Changes from the six-dimension schema

The corpus was first tagged on six dimensions, and the framework revises and extends them:

Two candidates were left out:

Until the next tagging run, Capability, Stance and the Prose scores are read from the current data. Scope, Horizon, Actor and Expression are charted from their six-dimension predecessors. The six new indices are not yet tagged.

Sources

Transcripts come from the public archives below. GINC adds transcripts of other public addresses where coverage is missing.