💬 NLP

Uncovering Temporal Framing in the News

This paper introduces a taxonomy and a multilingual dataset of 458 news articles annotated for temporal framing, demonstrating that supervised models effectively detect these persuasive rhetorical devices while outperforming zero-shot approaches.

Original authors: Tarek Mahmoud, Veronika Solopova, Premtim Sahitaj, Ariana Sahitaj, Max Upravitelev, Mervat Abassy, Hana Fatima Shaikh, Neda Foroutan, Vera Schmitt, Preslav Nakov

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tarek Mahmoud, Veronika Solopova, Premtim Sahitaj, Ariana Sahitaj, Max Upravitelev, Mervat Abassy, Hana Fatima Shaikh, Neda Foroutan, Vera Schmitt, Preslav Nakov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are reading a news article. Usually, when you see words like "yesterday," "next year," or "since 1990," your brain treats them like a calendar. You think, "Okay, this happened then, and that will happen later." It's just a timeline.

But this paper argues that news writers often use time not just as a calendar, but as a magic wand to change how you feel about a story. They aren't just telling you when something happened; they are using time to persuade you that something is urgent, nostalgic, or inevitable.

The researchers call this "Temporal Framing."

The Eight Magic Tricks (The Taxonomy)

The team created a "menu" of eight different ways writers use time to pull your emotional strings. Think of these as eight different lenses a writer can put on a camera to change the picture:

Primacy (The "First Mover" Effect): "We were the first to discover this!" It makes the writer sound like a pioneer or an authority.
Recency (The "Breaking News" Hype): "This happened just now!" It makes old facts seem irrelevant and the new story feel like the only thing that matters.
Urgency (The "Ticking Clock"): "We have only 48 hours!" It creates panic and forces you to act immediately.
Temporal Anchoring (The "Historical Anchor"): "We live in a post-9/11 world." It ties the current event to a famous past moment to borrow its emotional weight.
Nostalgia (The "Rose-Tinted Glasses"): "Remember the good old days?" It makes the past look perfect to make the present look bad.
Temporal Contrast (The "Then vs. Now"): "It was a booming city then, but a ghost town now." It highlights a dramatic change to tell a story of decline or progress.
Continuity (The "Never-Ending Story"): "This has been happening for decades." It makes a situation feel unchangeable, inevitable, or frustratingly stuck.
Skeptical (The "Cloudy Crystal Ball"): "This plan might collapse." It casts doubt on the future to make people cautious or afraid.

The Detective Work (The Dataset)

To study this, the researchers didn't just guess. They built a massive database of 458 news articles in English and German.

The Hunt: They read thousands of sentences and asked human experts to act like detectives.
The Task: The experts had to spot sentences where time was being used as a "persuasion tool" rather than just a fact.
The Result: They found over 2,000 sentences where time was being used as a rhetorical trick. They labeled them with the eight categories above.

The Robot Test (The Experiments)

Once they had this "training manual," they tried to teach computers to spot these tricks. They ran two types of tests:

The "Zero-Shot" Test (The Guessing Game): They asked big, smart AI models (like Llama and Qwen) to guess the frames without any special training.
- The Result: The AI was okay at guessing, but it often got confused. It would see a word like "now" and assume it was urgent, even if it wasn't. It was like a student who memorized a dictionary but didn't understand the context of a joke.
The "Supervised" Test (The Training Camp): They took those same AI models and fed them the 458 articles they had already labeled. They taught the models exactly what to look for.
- The Result: Huge improvement. The trained models became much better at spotting the tricks. They learned that "now" doesn't always mean "urgent" and that "since" doesn't always mean "nostalgic."

The Big Takeaway

The paper concludes that time is a language of persuasion.

It's learnable: Computers can learn to detect these rhetorical tricks if they are taught properly.
It's tricky: You can't just look for time words (like "yesterday" or "soon"). You have to understand the story the writer is telling. A sentence saying "Inflation rose in 2024" is just a fact. But "Inflation has been rising for years, proving the policy failed" is a frame.
Size isn't everything: Just making an AI bigger doesn't automatically make it better at this. It needs specific training to understand the difference between a calendar date and a rhetorical weapon.

In short, this paper gives us a new way to look at the news: not just at what happened, but at how the writer used the concept of time to make you feel a certain way about it.

Technical Summary: Uncovering Temporal Framing in the News

Problem Statement

While Natural Language Processing (NLP) has made significant strides in temporal expression extraction and temporal reasoning, these efforts predominantly treat time as an objective, descriptive property of events (e.g., predicting verb tense or ordering events chronologically). Current systems largely overlook the rhetorical and persuasive functions of temporal language. In news discourse, references to the past, present, and future often serve as framing devices that shape interpretation, legitimize authority, and influence decision-making, rather than simply reporting chronology. This gap limits the ability of computational systems to analyze persuasive rhetorical structures along the temporal dimension, particularly in domains like political communication where temporal justifications are highly influential.

Methodology

1. Taxonomy Construction

The authors propose a taxonomy of eight temporal frames grounded in social science research (psychology, communication, and political science). These frames are defined not by surface-level temporal expressions alone, but by their rhetorical function in structuring meaning:

Primacy: Significance attributed to being first in time.
Recency: Significance attributed to the most recent events.
Urgency: Emphasis on limited time or imminent consequences.
Temporal Anchoring: Framing discussion through the lens of past events.
Nostalgia: Invocation of a cherished past.
Temporal Contrast: Juxtaposition of "then" versus "now."
Continuity: Persistence across time (stability, stagnation, or momentum).
Skeptical: Casting doubt about the future.

The task is formalized as a sentence-level, multi-label classification problem: given a document $D$ and a sentence $s_i$ , the goal is to predict the set of temporal frames $P(F)$ expressed in $s_i$ .

2. Corpus Creation

The authors constructed a multilingual news corpus comprising 458 English and German articles from a diverse set of outlets (e.g., The Guardian, Fox News, Die Welt, Bild).

Selection: Articles were retrieved from GDELT (September 2024–September 2025). To increase the density of persuasive content, the authors used LLM-assisted tagging (GPT-4o) to identify opinionated articles, resulting in a final dataset where ~70% of documents were opinionated.
Annotation: A team of human annotators (3 English, 4 German) performed sentence-level annotation using the INCEpTION platform.
Guidelines: Annotators were instructed to distinguish between factual chronological reporting (neutral) and opinionated temporal framing (persuasive). Crucially, they were directed to ignore quoted or indirect speech, focusing only on the authorial voice.
Scale: The dataset contains over 2,000 temporally framed sentences and approximately 3,000 annotations derived from a corpus of over 20,000 sentences.
Quality: Inter-annotator agreement (Krippendorff's $\alpha$ ) reached 0.605 (English) and 0.720 (German) for binary detection, with moderate agreement on specific frames, consistent with prior fine-grained framing tasks.

3. Experimental Setup

The study evaluates temporal framing detection using two primary strategies:

Zero-Shot Classification: Using large language models (LLMs) including GPT-5.2, LLaMA-3 variants (8B, 70B), and Qwen3 variants (8B to 235B) without task-specific training.
Supervised Fine-Tuning: Fine-tuning encoder-based models (XLM-RoBERTa) and decoder-based LLMs (Qwen3-8B, LLaMA-3.1-8B) on the annotated corpus.
Input Strategy: Experiments utilized a sentence-only input, as adding neighboring context or full document context was found to make models more conservative and did not improve performance.

Key Results

1. Learnability of Temporal Framing

The experiments demonstrate that temporal framing is learnable at the sentence level, but it is not reliably captured by model scale alone in zero-shot settings.

Zero-Shot Performance: Zero-shot models exhibited high recall but low precision, indicating a tendency to over-predict frames. Performance improved with model size (e.g., Qwen3-235B outperformed Qwen3-8B), but gains were modest and quickly saturated. GPT-5.2 achieved the best zero-shot binary F1 of 0.45.
Supervised Performance: Supervised fine-tuning yielded substantial improvements.
- XLM-RoBERTa (270M parameters) achieved a binary F1 of 0.51 and Micro-F1 of 0.37, outperforming all zero-shot baselines despite its smaller size.
- Fine-tuned LLMs achieved the strongest results. Qwen3-8B reached a binary F1 of 0.57 and Micro-F1 of 0.44 (a 27% relative improvement over the best zero-shot result). LLaMA-3.1-8B showed a 125% relative improvement over its zero-shot counterpart.

2. Frame-Specific and Cross-Lingual Analysis

Frame Variability: Frames like Continuity, Temporal Contrast, and Urgency were detected more reliably due to their prevalence and recurring lexical cues. Rare frames, particularly Nostalgia, remained difficult to detect due to data sparsity and reliance on implicit evaluative language rather than explicit temporal markers.
Language Differences: Performance was generally higher on English data than German data. The authors attribute this to the German subset containing fewer framed sentences and fewer positive instances per frame, rather than a failure of cross-lingual generalization.

Significance and Contributions

The paper makes four primary contributions:

Conceptualization: It casts temporal framing as a distinct, sentence-level, multi-label task that captures the rhetorical use of time beyond surface expressions.
Taxonomy: It introduces an eight-frame taxonomy grounded in social science theory, bridging the gap between computational linguistics and rhetorical analysis.
Resource: It releases a multilingual, sentence-level annotated news dataset (English and German) to support future research on temporal framing.
Empirical Baselines: It provides computational baselines demonstrating that while large language models possess implicit knowledge of time, task-specific supervision is required to effectively learn the systematic but subtle patterns of temporal framing.

The authors conclude that temporal framing occupies a distinct space between surface-level temporal expression and higher-level narrative structure. While models can learn to detect these patterns with appropriate training signals, the task requires interpreting implicit rhetorical cues that extend beyond explicit temporal markers, highlighting the need for richer discourse representations in future work.