Misinformation Detection

A natural language inference task that classifies the attitude of a text segment toward a target claim. Unlike sentiment analysis, stance detection determines if a headline agrees, disagrees, discusses, or is unrelated to a proposition.

• Uses premise-hypothesis pair modeling
• Often deployed as a first-pass filter before full fact-checking
• Example: Given the claim 'Vaccines cause autism,' a headline stating 'Study debunks vaccine-autism link' is classified as disagreeing
• Architectures typically fine-tune BART or T5 models on the FNC-1 dataset

The process of breaking a complex, multi-faceted statement into atomic, verifiable sub-claims. A single sentence like 'The CEO resigned after the stock crashed 40%' contains two distinct factual assertions that must be independently verified.

• Uses semantic role labeling to identify predicates and arguments
• Enables granular matching against knowledge bases
• Prevents partial truth from granting full credibility
• Critical for handling check-worthy political speech and earnings calls

Extracting psycholinguistic markers that correlate with deceptive content. Misinformation often exhibits distinct patterns in pronoun usage, certainty adverbs, and concreteness.

• High use of first-person plural ('we,' 'us') can signal in-group manipulation
• Hyperbolic adjectives ('devastating,' 'unprecedented') appear at elevated rates
• Readability scores and syntactic complexity serve as weak signals
• Tools like LIWC (Linguistic Inquiry and Word Count) provide standardized lexicons for feature extraction

The retrieval-augmented pipeline that searches a trusted document corpus for evidence supporting or refuting a detected claim. This transforms fact-checking from a closed-book generation task into an open-book verification task.

• Queries are constructed from claim entities and relations
• Dense retrieval over Wikipedia, PubMed, or proprietary knowledge graphs
• Returns ranked passages with attribution metadata
• The retriever must balance precision (finding relevant evidence) with recall (not missing critical refutations)

Modeling the temporal and structural dynamics of how a piece of content spreads through a network. Misinformation often exhibits distinct cascades compared to factual news.

• Burstiness: False news often spikes rapidly via coordinated shares
• Depth vs. Breadth: Rumor cascades tend to be deeper and narrower
• User Account Features: Analyzes account age, follower ratios, and posting frequency
• Graph neural networks model the retweet graph to classify propagation trees as organic or orchestrated

Extending detection beyond text to images and video. Modern misinformation often uses authentic media in a false context, a tactic known as out-of-context misuse.

• Reverse image search to find prior instances of a photo
• Error Level Analysis (ELA) to detect splicing and manipulation
• Semantic inconsistency detection between image content and caption claims
• Deepfake detection using physiological signals like heartbeat patterns in facial blood flow

A verification technique that boosts the confidence score of a factual claim when multiple independent, authoritative sources corroborate the same information. This approach mitigates single-source bias and reduces the risk of amplifying an isolated error.

• Source Independence: Requires that corroborating outlets do not share a parent entity or syndication feed.
• Temporal Analysis: Weighs agreement higher when sources confirm a fact simultaneously without cross-citation.

A sub-task of Natural Language Processing that automatically determines the perspective of a text towards a target claim, typically categorized as agree, disagree, discuss, or unrelated. Unlike fact-checking, stance detection analyzes viewpoint without needing to verify objective truth.

• Feature Engineering: Relies on linguistic markers, sentiment, and dependency parsing.
• Application: Used to identify echo chambers and map the spread of misleading narratives across social graphs.

The process of documenting the origin, custody, and transformation history of a piece of information to establish its authenticity and chain of attribution. In misinformation detection, provenance tracking helps identify if a video or image has been decontextualized or synthetically altered.

• Metadata Inspection: Analyzing EXIF data and creation timestamps.
• Reverse Image Search: Identifying prior instances of a visual asset to detect recirculation of old media in a new, false context.

The computational analysis of text to identify subjective language, framing, or one-sided argumentation that indicates a lack of editorial neutrality. Algorithms scan for loaded terminology and selective omission of facts.

• Linguistic Markers: Identifies hyperpartisan phrases, emotional polarization, and ad hominem attacks.
• Structural Bias: Detects imbalance by measuring the word count and prominence given to opposing viewpoints within a single document.

A probabilistic framework that updates the trustworthiness score of a source by combining prior beliefs with new evidence of content accuracy or deception. This model adapts dynamically as a publisher's track record evolves.

• Prior Probability: The initial trust score based on domain history or expert seed lists.
• Likelihood Ratio: Adjusted upward for verified true claims and penalized exponentially for confirmed falsehoods.