This model is a fine-tuned version of FacebookAI/xlm-roberta-large trained on a custom human-labelled dataset of social media posts published by major news media outlets in six countries: U.S., U.K., Ireland, Poland, France and Spain between 2020 and 2024.

The purpose of the model is to discriminate between different kinds of valence in news articles, headlines, and posts. To define valence, we followed the conceptual work by Lengauer et al. (2011), who proposed definitions and measurements of valence in news. Our operationalization and classifier define valence by distinguishing between “the mere dissemination of negative news” (exogenous valence coming into the news from outside, that is, from the topic itself) and “endogenous valence imposed on news by journalists through their usage of language”. More concretely, it assesses two aspects separately:

Research purposes, in particular selection of texts from large diverse corpora and/or calculation of statistics in groups (i.e. for negative and non-negative content). The design and conceptualization of this model was tailored for a specific research project and may not be relevant in other contexts.

The model should work well for the languages it was fined-tuned on. However, since it is based on a multilingual backbone it may also work relatively well for other languages. That said, in such cases a noticeable drop in performance is expected.

Cannot be shared for legal reasons. The scores obtained on a validation hold-out subset of the dataset were:

The model uses custom processing and pipeline logic, so has to be initialized in a slightly more verbose way than standard models. For the same reason, it has to be loaded with the trust_remote_code=True flag. Last but not least, this model has two separate classification heads, and thus to be used as a transformers pipeline it must select a custom text-multi-classification task.

Now, the pipelines is ready to be applied to text data in the standard way.