Emotion contains essential information in the translation/interpretation process, contributing to evaluation and decision-making tasks, since emotions inform cognitive processes and people from different cultures react and behave differently (Caruso and Salovey, 2004). Furthermore, a person (speaker) who may be diligent in communication and interactions in one’s own culture may not be equally diligent in respective cases concerning situations and people from other cultural backgrounds (Davaei et. al. 2022).

Despite remarkable achievements in recent research, the creation and maintenance of better world models in monolingual and multilingual applications with Large Language Models (LLMs) involving human behavior and emotion remains a challenge. The efficient integration of in-depth world knowledge concerning language and culture-specific and/or domain-specific information in AI and Natural Language Processing applications is an essential requirement for the creation of robust world models.

The proposed “human-in-the-loop” approach in the modelling of world knowledge – in this case, the translator’s/interpreter’s knowledge -  for behavior and emotion functioning as a “road-map” of key-parameters in the language(s) and/or domain(s) concerned contributes to the creation of world models for monolingual and multilingual AI applications and training data, including LLMs.  World models reflecting human behavior and human values in the real world can be based on world knowledge - Collective Intelligence –integrated in analysis and processing strategies from text types such as political and journalistic texts, containing socio-cultural elements and involving essential information for the translation/interpretation process and for understanding human behavior, emotion and, ultimately, the human mind.  

This complex information can be processed in respect to a “three-level” framework with distinct, identifiable features (“User-Group”, “Pragmatic”, “Prosodic-Paralinguistic” Levels), serving  as a basis for proposed “plug-in” customizations for language/culture or domain-specific human behavior and emotions – with their complexity factors (“Subtlety”, “Intensity”, “Manner” and “Interference”). These proposed “plug-in” customizations are in the form of (a) customized Knowledge Graphs – subgraphs, (b) additional resources combined with LLMs or (c) upgraded rule-based approaches and seed data.

The proposed processing strategies include empirical data (and characteristic cases- examples in English, German and Greek, especially transcribed spoken political and journalistic texts) collected/ processed with the collaboration of professional journalists for the Journalism Computational Linguistics Research Lab – European Communication Institute (ECI), Donau Universität Krems (DUK) Austria.