Social media is one amongst the most available news sources these days for a lot of folks around the globe, thankful for their low value, quick access, and speedy dissemination. However, this comes at the value of questionable traits and vital risks for exposure to ‘fake news’ that are purposely written to mislead the readers. Such information has the potential of touching vox populi or voice of the public and providing a chance for malicious parties to control the outcome of public events, like elections.
With unreliable words, people can get affected by fake news very easily and will share them without any fact-checking. For instance, throughout the 2016 United States presidential election, various styles of pretended news concerning the candidates widely, and unfold through each official journalism and on-line social networks.
With a thorough investigation of fake news data, lots of useful and explicit features can be identified from both the text words and images used in the fake news. Besides these specific benefits, there are also some hidden patterns that exist in the words and images used in fake news, which can be captured with a set of latent features extracted via the multiple convolutional layers in the model.
Convolutional Neural Network (Deep Model for fake news detection)
A faux news detection system aims to help users in detecting and filtering out varieties of potentially deceptive news. The prediction of the possibilities that a specific news item is designedly deceptive is predicated on the analysis of previously seen truthful and deceptive news.
We discuss the three types of fake news, each in contrast to genuine serious reporting:
1. Serious Fabrications: They present a wide spectrum of unverified new and uses eye-catching headlines, exaggerations, scandal-mongering, or sensationalism to increase traffic or profits. These fabrications specifically emphasize topics such as sensational crime stories, astrology, gossip columns about celebrities, and junk food news.
2. Large-Scale Hoaxes: Hoaxing is another style of deliberate fabrication or falsification within the thought or social media. Attempts to deceive audiences masquerade as news, and maybe picked up and mistakenly validated by traditional news outlets.
3. Humorous Fakes: We distinguish serious fabricated news from humorous ones. If readers are aware of the humorous intent, they may no longer be predisposed to take the information at face value. Technology will establish humor and conspicuously show originating sources (e.g., The Onion) to alert users, particularly in decontextualized news aggregators/platforms.
Neural Network-based Approaches (Convolutional Neural Network) for deceptive news detection
Deep learning models are widely used in both the academic community and industry, in computer vision and speech recognition, the state-of-art methods, etc, are almost deep neural networks. Researchers additionally noticed that CNN is effective on several IP tasks, for instance, semantic parsing, sentence modeling, and other traditional NLP tasks.
This kind of network works by applying a series of ﬁlters to their input. These ﬁlters are N−dimensional matrices which are slid (convoluted) over the input, which means that a series of matrix multiplications are executed in chunks over that input. After training the network, the ﬁlters produce activations (known as feature maps) where certain patterns are detected (for example, in images, these are borders, ﬁgures, patterns, etc.)
We’ll be going with this convolutional neural network model for our fake news detection.
“This is going to be the beguiling part of the blog for all the data enthusiasts.”
This is the data used to train the deep model. In 2017, William created a new benchmark dataset called LAIR which collected 12.8K manually labeled short statements that are labeled in different contacts from Politifact, it provides detailed analysis reports and links to source documents for each case. William works on the LAIR dataset with many techniques such as Logistic regression, Support vector machines, Bidirectional-LSTM and CNN models for Deep learning in which CNN results are the best.
The LIAR dataset includes 12,836 short statements labeled for truthfulness, subject, context/venue, speaker, state, party, and prior history. With this size and time span of ten years, LIAR cases are collected in a more natural context, such as political debate, TV ads, Facebook posts, tweets, interviews, news releases, etc. In each case, the labeler provides a lengthy analysis report to the ground each judgment.
We can find out some interesting facts about fake news such as
Firstly much fake news has no titles. This fake news is widely spread as the tweet with a few keywords and hyperlinks of the news on social networks.
Secondly, there are a lot of capital characters in faux news, the purpose is to draw the readers’ attention, while the real news contains less capital letters, which is written in a standard format.
Thirdly, the real news contains more detailed descriptions, for example, names (Jeb Bush, Mitch McConnell, etc.), and motion verbs (left, claim, discussion and poll, etc.)
We investigate the text and image information from various perspectives, such as computational linguistics, sentiment analysis, psychological analysis, etc.
1. Cognitive Perspective: From the cognitive perspective, we investigate the exclusive words (e.g. ‘but’, ‘without’, ‘however’) and negations (e.g. ‘no’, ‘not’ ) used in the news. Truth tellers use negations more frequently.
2. Psychology Perspective: Deceptive people often use language that minimizes references to themselves. A person who is lying tends not to use “we and “I”, i.e. tends not to use personal pronouns. Instead of saying “I didn't take your book”, a liar person might say “That’s not the kind of thing that anyone with integrity would do”.
On average, fake news has fewer first-person pronouns (e.g., you, yours) and third-person pronouns (e.g., he, she, it) and also tallied up.
3. Lexical diversity: Lexical diversity is a measure of how many different words that are used in a text, while lexical density provides a measure of the proportion of lexical items (i.e. nouns, verbs, adjectives, and adverbs) in the text. The rich news has more diversity.
4. Sentiment Analysis: The sentiment, in real and fake news, is totally different. For real news, they are more positive than negative ones. The reason is that deceivers may feel guilty or they are not confident about the topic. Under the tension and guilt, the deceivers may have more negative emotions.
The standard deviation of fake news on negative sentiment is also larger than that of real news, which indicates that some of the fake news has a very strong negative sentiment.
Beyond the text information, images in fake news are also different from those in real news. As shown in the figure, cartoons, irrelevant images (mismatch of text and image, no face in political news) and altered low-resolution images are frequently observed in fake news.
A glimpse of fake and real image analysis of any random image published in a local newspaper
Model (Convolutional Neural Network)
Besides the explicit features, we tend to innovatively utilize two parallel CNNs to extract latent features from each textual and visual info. And then, specific and latent features are projected into constant feature areas to create new representations of texts and pictures.
At last, we fuse textual and visual representations together for fake news detection. The overall model contains two major branches, i.e. text branch and image branch. For each branch, taking textual or visual data as inputs, explicit and latent features are extracted for final prediction.
1. Text Branch: For the text branch, we utilize two types of features: Textual Explicit Features XT e and Textual Latent Features XT l. The textual explicit features are derived from the statistics of the news text as we mentioned in the data analysis part, such as the length of the news, the number of sentences, question marks, exclamations, and capital letters.
Latent features are 'hidden' features to distinguish them from observed features. Latent features are computed from observed features using matrix factorization. With the convolutional approach, the neural network can produce local features around each word of the adjacent word and then combine them using a max operation to create fixed-sized word-level embeddings. Therefore, CNN is employed to model textual latent features for fake news detection.
Image and text analysis of news using CNN
2. Image Branch: Similar to the text branch, we use two types of features: Visual Explicit Features XI e and Visual Latent Features XI l. In order to obtain the visual explicit features, we firstly extract the resolution of an image and the number of faces in the image to form a feature vector. Then, we transform the vector into our visual explicit feature with a fully connected layer.
Finally, we utilize CNN to combine the explicit and latent features of text and image information into a unified feature space, and then use the learned features to identify the fake news.
Regarding the future improvement of these models, firstly, it is mandatory to collect more data, especially from a recent period of time. This is also proposed by the researches who compiled the TI-CNN dataset, as the news there is mostly obtained during the US electoral campaign. In order to accomplish the above proposal, a system to automatically collect quality news should be developed. Finally, with the aim that these models can be used by the people, some method of serving them to users is also necessary (integration with social networks, browser extensions, mobile apps).
The spread of fake news has raised concerns all over the world recently. These fake political news may have severe consequences. The identification of fake news grows in importance. In recent years, deception detection in online reviews & fake news has an important role in business analytics, law enforcement, national security, political due to the potential impact fake reviews can have on consumer behavior and purchasing decisions. Researchers used deep learning with the large dataset to increase in learning and thus get the best results by using word embedding for extracting features or cues that distinguish relations between words in syntactic and semantic form.
The experience gained during the development of these models allows us to state that the use of deep learning models for this task can be potentially beneficial for a wide range of actors, from social network companies to the final user in order to mitigate the increasing deceptions on the Internet. For more blogs in Analytics and new technologies do read Analytics Steps.