Amid the campaign noise, TOI used a large language model (LLM) with a fixed prompt to analyse speeches by three key leaders from April 14–20 in the TN 2026 election, processing a 50,000+ word corpus.The model broke the text into “tokens” (words) and detected repeated phrases.A separate prompt mapped the tone of the leaders’ speeches by tagging segments into five categories: political, administrative, emotional, populist, and ideological, covering what is said and how it is conveyed.Next, the model was instructed to classify content into predefined issue topics, including welfare/ promises, governance, corruption, price rise/taxes/economics, law and order, political attacks, state rights, and ideology. We limited the topics to six for a common baseline; adding more splits similar ideas across labels and reduces classification accuracy.Finally, the prompt asked the model to build a “speech-o-meter” to score each leader’s speeches on effectiveness using word use, repetition, topics, and tone. A custom analytical framework ensured comparability across leaders. The model assigned weights out of 10: issue dominance (3), recall value (2), repetition efficiency (1.5), emotional connect (1.5), lexical clarity (1), and tone balance (1), which were aggregated to produce the final score.
