poltextlab / xlm-roberta-large-illframes-covid

xlm-roberta-large-illframes-covid

Illframes Project - Covid

(illiberal policy frames) The ILLFRAMES Project's goal is to identify illiberal policy frames in texts. The codes are mutually exclusive and unequivocal. The codebook currently only covers the policy domain of migration and covid, but other policy domains such as climate, public health, and democracy will be added in the near future.

The project's covid domain focuses on identifying different types of critical discourse surrounding a public health crisis and government responses. The labels capture various narratives, including skepticism toward the severity of the crisis (e.g., exaggeration by experts), conspiracy theories about elite control and surveillance, economic and social harm caused by restrictions, and concerns over media bias and global power shifts. The classification also distinguishes between criticisms of political leaders, expert discourses, media coverage, and the broader societal consequences of public health measures.

Codebook:

• 310: This label includes texts containing critical, sceptical, conspiratorial statements regarding a public health situation and the government's handling thereof which cannot be put in any other categories. Criticism of (lack of ) government attention and spending in relation to the public health crisis belong to this category, if it does not specify its causes or consequences.
• 311: This label depicts the severity of the public health emergency as exaggerated by political leaders and experts. It attributes malicious intent (e.g., power grab), over-cautiousness, and/or incompetence to representations of the public health situation as extraordinary, critical, serious, and/or crisis-like. Portrayals of health policy experts and scientific advisors as untrustworthy, unreliable, out of touch with reality or similar belong to this category. In contrast with the elite control frame, the threat scepticism frame questions whether the spreading illness is truly a problem. In contrast with the truth frame, the threat scepticism frame focuses on political and expert discourses regarding the seriousness of the public health situation, not on the media treatment of the issue.
• 312: This label depicts the public health emergency as the result of a deliberate ploy by political/social/economic elites to control society. It criticises social monitoring and surveillence methods (e.g., tracking movement via a phone app). It suggests that restrictions are used to take away personal freedoms (e.g., the right to assemble and the right to protest) and curb democratic oversight. Accusations of elites with the fabrication of patogens and viruses to weaken social resistence against them belong to this category. In contrast with the threat scepticism frame, the elite control frame does NOT diminish the severity of the public health situation.
• 313: This label depicts government restrictions to tackle the public health crisis as destroying small businesses and undermining the proper functioning of the economy; thus (in)directly resulting in more significant social harm. It criticises the government proposals and action for leading to (further) indebtedness, causing high inflation, hruting economic growth and compromising competitiveness. Criticisms of social distancing, vaccination and mask mandates as hindering the functioning of small and medium-sized entreprises belong to this category.
• 314: This label depicts government measures (e.g., obligatory vaccination, mask-wearing, social distancing, obligatory testing to attend social events) to tackle the public health crisis as infringing on personal freedoms and/or human rights. It advocates for health privacy and unristricted medical choice, including the option not to medicate. It suggests that public health measures, testing, medication and/or vaccination are unreliable and/or medically harmful, for example, by increasing substance abuse or hindering the treatment of other illnesses. Citing conspiracy theories (e.g., vaccines are used to inject chips) or casting medication as more dangerous than the illness to justify opposition also belong to this category. Criticisms of people facing repercussions for not taking vaccination, medication, tests, etc., belong to this category.
• 315: This label depicts the media and journalists as servants of the elites and complicit in public health-related fearmongering. It casts media coverage of the public health crisis as biased and hiding the truth to spread fear and ensure social compliance with government orders. Criticising the media for not giving adequate coverage to discussion of the "woke virus" and illness-related conspiracy theories (e.g., mind control chips in vaccines) belong to this category. In contrast with the threat scepticism frame, the truth frame focuces on the media and public intellectuals, NOT on politicians or experts.
• 399: None of them

How to use the model

from transformers import AutoTokenizer, pipeline

tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-large")
pipe = pipeline(
    model="poltextlab/xlm-roberta-large-illframes-covid",
    task="text-classification",
    tokenizer=tokenizer,
    use_fast=False,
    truncation=True,
    max_length=512,
    token="<your_hf_read_only_token>"
)

text = "We will place an immediate 6-month halt on the finance driven closure of beds and wards, and set up an independent audit of needs and facilities."
pipe(text)

Classification Report

Overall Performance:

• Accuracy: 61%
• Macro Avg: Precision: 0.33, Recall: 0.26, F1-score: 0.27
• Weighted Avg: Precision: 0.61, Recall: 0.61, F1-score: 0.59

Per-Class Metrics:

Inference platform

This model is used by the CAP Babel Machine , an open-source and free natural language processing tool, designed to simplify and speed up projects for comparative research.

Cooperation

Model performance can be significantly improved by extending our training sets. We appreciate every submission of CAP-coded corpora (of any domain and language) at poltextlab{at}poltextlab{dot}com or by using the CAP Babel Machine .

Debugging and issues

This architecture uses the sentencepiece tokenizer. In order to run the model before transformers==4.27 you need to install it manually.

If you encounter a RuntimeError when loading the model using the from_pretrained() method, adding ignore_mismatched_sizes=True should solve the issue.