Ensemble Techniques for Robust Fake News Detection: Integrating Transformers, Natural Language Processing, and Machine Learning

. 2024 Sep 19;24(18):6062. doi: 10.3390/s24186062

Algorithm 1 Multi-modal Disinformation Detection

1:
Input: Raw text data T, Image data I
2:
Output: Comprehensive representation $C$
3:
Textual Feature Extraction:
4:
Tokenize text: $T_{t} = tokenize (T)$
5:
Normalize words: $T_{n} = normalize (T_{t})$
6:
Replace emojis: $T_{e} = replace_emojis (T_{n})$
7:
Shorten sentences: $T_{s} = shorten (T_{e})$
8:
Extract BERT embeddings:
$E = BERT (T_{s}) where E = [h_{- 4}, h_{- 3}, h_{- 2}, h_{- 1}]$
9:
Combine embeddings: $T_{f} = combine (E)$
10:
Visual Feature Extraction:
11:
Pre-trained ResNet V2 model:
$I_{1} = ResNet (I)$
12:
Fully connected layers:
$I_{f} = FC (I_{1})$
13:
Process visual representation:
$I_{m} = process (I_{f}) where d_{I} = 16$
14:
Attention Mechanism:
15:
Apply attention:
$A_{T \to I} = Attention (T_{f}, I_{f})$

$A_{I \to T} = Attention (I_{f}, T_{f})$

$A_{I \to I} = Attention (I_{f}, I_{f})$
16:
Fully connected layers with normalization:
$R_{T \to I} = FC (A_{T \to I}) + T_{f}$

$R_{I \to T} = FC (A_{I \to T}) + I_{f}$

$R_{I \to I} = FC (A_{I \to I}) + I_{f}$
17:
Final Processing:
18:
Compress and combine features:
$R_{I \to I}^{'} = FC (R_{I \to I})$
19:
Fully connected layer with 32 neural units:
$C = {FC}_{32} ([T_{f}, I_{f}, R_{T \to I}, R_{I \to T}, R_{I \to I}^{'}])$