Source-linked AI summary
An Emotional Analysis of False Information in Social Media and News Articles
Bilal Ghanem, Paolo Rosso, Francisco Rangel
TL;DR
False information can manipulate readers, motivating analysis of whether emotional language distinguishes false from real news across social media and online articles. The paper combines emotional features from multiple lexicons with an LSTM-based network and evaluates it across false-information types. Results show type-specific emotional patterns, a key role for emotions in deception, and near-96% F1 on a clickbait dataset.
Problem
False information can manipulate readers’ opinions, but the paper investigates how emotional language differs between false and real news across sources and false-information types.
Method
The paper combines content and emotional features from multiple lexicons in an emotionally infused LSTM network for false-information detection.
Results
Near 96% F1 was achieved on a clickbait dataset, while emotional features improved performance across three datasets and false-information types showed distinct emotional patterns.
Takeaways & Limitations
Emotions are useful signals for detecting false information, with their roles varying by type and similar overall importance distributions across Twitter and news articles.
Takeaways & Limitations
The study uses different validation scenarios because cross-validation was computationally expensive for tuning the deep neural network.
Abstract
from arXiv · showhide
Fake news is risky since it has been created to manipulate the readers' opinions and beliefs. In this work, we compared the language of false news to the real one of real news from an emotional perspective, considering a set of false information types (propaganda, hoax, clickbait, and satire) from social media and online news articles sources. Our experiments showed that false information has different emotional patterns in each of its types, and emotions play a key role in deceiving the reader. Based on that, we proposed a LSTM neural network model that is emotionally-infused to detect false news.
1 INTRODUCTION
The paper frames false information as a growing threat across social media and online news, and investigates whether emotional language can distinguish its types and support detection. It proposes combining emotional analysis with a deep neural network.
- Scope: The study analyzes hoaxes, propagandas, clickbaits, and satires across Twitter and online news articles.For news articles, the analysis focuses on their beginnings because article length could bias emotional analysis.
- Motivation: False information can be intentionally designed to influence readers’ opinions, while satire is presented without intent to deceive.The paper distinguishes misinformation, such as satire, from disinformation, including hoaxes, propaganda, and clickbait.
- Motivation: The paper argues that emotional patterns in false news may help explain how writers trigger readers’ emotions and contribute to fake-news detection.Its motivation includes the growing spread of suspicious news and organized efforts to affect public opinion.
- Research questions: The research asks whether emotional features help detect false information and whether emotion distributions differ across sources and false-information types.It also examines which emotions statistically distinguish false from real news.
- Contributions: The proposed approach combines emotional information from documents with a deep neural network and compares it with baselines.The article also presents a comprehensive affective analysis of two datasets from social media and online news.
2 RELATED WORK
Related work has examined false information through linguistic, credibility, and claim-based signals, but the paper describes prior approaches to false-information analysis as relatively limited. Existing findings include differences in bias, subjectivity, harmful language, moral framing, and user-related credibility features.
- Language analysis: Prior linguistic studies found that real tweets contain fewer bias markers, hedges, subjective terms, and harmful words than false tweets.They also reported differences in moral targeting among propaganda, satire, hoax, and clickbait news.
- Language analysis: Earlier work reported that propaganda targets morals more than satires and hoaxes but less than clickbaits.Satirical news was associated with more loyalty and fewer betrayal morals than propaganda.
- Credibility detection: Credibility models have used message-, user-, and propagation-based features to distinguish truthful from deceitful tweets.User information such as age, follower counts, and status counts helped discriminate the classes most strongly in one study.
- Claim-based methods: Other approaches have focused on estimating the credibility of news claims by retrieving supporting information.The paper positions these efforts within a relatively small body of previous false-information analysis.
3 EMOTIONALLY-INFUSED MODEL
The Emotionally-Infused Network combines content representations and emotional features to predict false-information types. It uses multiple emotion lexicons, an LSTM with attention for text, and a second branch for emotion vectors before classification.
- Emotional Lexicons: The model uses EmoSenticNet, EmoLex, SentiSense, LIWC, and Empath to broaden emotional-word coverage and represent 17 emotions.These resources draw on different psychological emotion models and category sets.
- Content branch: The content branch converts word sequences into embeddings, processes them with an LSTM, and applies attention to emphasize particular words.The attention layer assigns class-focused weights to word-vector states.
- Content branch: The architecture includes attention because the model is designed to focus on particular words rather than weighting all word states equally.An experiment without attention produced a lower result.
- Emotional branch: The emotional branch represents an input document with emotion vectors generated from five lexicons.The resulting lexicon-specific vectors are combined into a single representation.
- Classification: The two branches are combined and passed through dense layers and a final Softmax function to predict output classes.The LSTM receives both content-based word embeddings and emotional features as network inputs.
4 EVALUATION FRAMEWORK
The evaluation uses publicly available data from news articles and Twitter to study emotional signals across real and false information. It compares the emotionally infused model with emotion-only and conventional neural and machine-learning baselines.
- Datasets: The evaluation relies on two publicly available datasets and suspicious Twitter-account lists because annotated false-information data are limited.The datasets support emotional analysis across different news sources.
- News articles: The news-article data combine trusted articles from English Gigaword with false articles from seven unreliable sites.The false-news collection includes satires, hoaxes, and propagandas, while clickbaits come from a separate available dataset.
- Twitter: The Twitter data include the last 1,000 tweets collected from accounts associated with each false-information type and real news.The real-news accounts combine a prior list with 32 additional Twitter accounts.
- Baselines: An emotion-only model is compared with Majority Class and Random selection baselines to test whether emotional features independently detect false news.The evaluation therefore assesses both the combined network and emotion-only information.
- Baselines: The emotionally infused model is compared with BOW-SVM, word-embedding Logistic Regression, and an otherwise identical LSTM without emotional features.These baselines isolate conventional lexical, embedding-based, and architecture-matched contributions.
5 EXPERIMENTS AND RESULTS
Experiments evaluated emotion-based and emotionally infused models across Twitter, news articles, and clickbait data. Emotional features detected false information, with EIN outperforming baselines and separating several false-information classes.
- Experimental setup: 20% of each dataset was reserved for testing, while 10-fold cross-validation selected and tuned the classical classifiers.Random Forest was chosen for both datasets because it achieved the best results among the tested classifiers.
- Emotion-based model: Emotional features clearly detected false news against baselines, performing better on news articles than on tweets.Class-level true-positive ratios were also examined against the random baseline.
- Emotion-based model: Clickbait achieved the highest true-positive ratio, while hoax detection exceeded the random baseline by only 4%.The news-articles dataset was balanced across false-information classes, whereas Twitter was biased toward propaganda.
- Emotionally-Infused Model: EIN combines emotional features with a neural document representation using dataset-specific model parameters, dropout, and categorical cross-entropy.For neural training, a separate validation split replaced cross-validation to reduce computational cost.
- Emotionally-Infused Model: EIN outperformed baselines by around 2% on Twitter and 7% on news articles, with the largest margin on news articles.The reported EIN results include 79.43% for news articles and 59.70% for Twitter in comparisons with LSTM.
- Emotionally-Infused Model: Document projections showed clickbait separating quickly, while later training separated false-information classes from real news with limited overlap among some classes.Residual overlap was reported between satire and hoax, and among a few propaganda records.
- EIN as Clickbaits Detector: On the clickbait dataset, EIN outperformed both the LSTM baseline and the Stop_Clickbait detector in a binary-classification experiment.The experiment used binary cross-entropy and replaced Softmax with a Sigmoid output.
6 DISCUSSION
Emotional features distinguish false from real news across Twitter and news articles, but their usefulness and distributions vary by source and false-information type.
- Emotional features improve false-news detection, especially for news articles, while Twitter detection is harder because informal language and typos reduce embedding coverage.Twitter also contains more sexual and insulting words than news articles, with normalized sexual-word means of 0.00319 versus 0.0024 and insulting-word means of 0.0027 versus 0.0017.
- The study examines emotional contributions across Twitter and news articles and asks whether emotion importance distributions are similar between the two sources.
- Information Gain was used to rank emotions that discriminate real news from false news in Twitter and news-article datasets.
- Eight emotions, including joy, negative emotion, disgust, trust, and surprise, differ significantly between real and suspicious news in both datasets.The analysis used t-tests across real and false news in a binary classification task.
- Clickbaits most strongly express surprise and negative emotion, whereas hoaxes lack a simple distinctive emotional pattern because they imitate truthful presentation.The clickbait pattern is reported across both datasets despite language differences.
7 CONCLUSIONS AND FUTURE WORK
The paper presents an emotionally infused deep-learning network for detecting false information in Twitter and news articles. Experiments find that emotions improve detection, with clickbait performance near 96% F1 and distinct emotional roles across false-information types.
- The emotionally infused network uses emotional features to identify false information in Twitter and news articles, with performance compared against an LSTM and other baselines.
- Near 96% F1 was achieved on a clickbait dataset, producing superior results to the reported state of the art.
- Emotional features boosted the EIN model across three datasets and helped detect false information in Twitter despite its sexual and insulting language.
- Emotions had similar importance distributions across Twitter and news articles, differed significantly between real and false news, and played different roles across false-information types.
- Future work proposes a clickbait-detection browser extension and analysis of how emotions flow through articles of each false-information type.