USING A SOURCE-LEVEL MANIPULATIVE SATURATION INDEX IN MULTI-CRITERIA SELECTION OF NEWS MESSAGES

Abstract

News selection systems that rank messages by engagement tend to amplify manipulative content, and manipulativeness may be a stable characteristic of the editorial practice of particular sources. Existing approaches typically treat these factors separately: they either penalize a manipulative message equally regardless of its source, or demote a source as a whole together with its non-manipulative messages. The article aims to investigate additive and multiplicative ways of accounting for the manipulative saturation of the source in multi-criteria selection of news messages and to compare them on real data. Message manipulativeness is estimated by an LLM-based multi-label classifier of ten manipulation techniques. The manipulative saturation index of a source is defined as the mean manipulativeness of the source’s messages in a time window and is validated as a measurement instrument. In the additive form, the index enters the objective function as a separate term and demotes all messages of the source; in the multiplicative form, it multiplies the penalty for message manipulativeness and does not demote non-manipulative messages. Both forms are compared with the penalty without the index and with each other at equal relevance cost on a frozen corpus of 125,987 messages from 34 sources over 182 days for 2,184 topical queries. The penalty on the manipulativeness of the message itself reduces the mean manipulativeness of shown messages 5.4-fold at a cost of 1.09 % of relevance. At equal cost, the additive form shifts positions away from saturated sources more strongly but also displaces their non-manipulative messages; the multiplicative form barely affects non-manipulative messages but reduces the share of positions held by saturated sources more moderately: by 15.6 % versus 62.8 % for the additive form, relative to the penalty without the index at the same cost.