In this paper, we present the Logically submissions to De-Factify 2 challenge (DE-FACTIFY 2023) on task 1 of Multi-Modal Fact Checking. We describe our submission to this challenge including explored evidence retrieval and selection techniques, pre-trained cross-modal and unimodal models, and a cross-modal veracity model based on the well established Transformer Encoder (TE) architecture which heavily relies on the concept of self-attention. Exploratory analysis is also conducted on the Factify 2 data set that uncovers the salient multi-modal patterns and hypothesis motivating the architecture proposed in this work. A series of preliminary experiments were done to investigate and benchmark diferent pre-trained embedding models, evidence retrieval settings and thresholds. The final system, a standard two-stage evidence based veracity detection system, yielded a weighted average F1 score of 0.79 on both the validation set and final blind test set of task 1, which achieved 3rd place with a small margin to the top performing systems on the leaderboard among 9 participants.
Misinformation and fake news can spread rapidly and cause harm at various levels. One way to protect ourselves from these negative impacts is through fact-checking and debunking false information with evidence-based reporting. However, this process can be resource-intensive and time-consuming. To address this issue, researchers have developed automated fact-checking systems using deep learning techniques, which can handle tasks such as claim detection, claim matching, evidence retrieval, and veracity prediction using natural language processing techniques on textual content. While there has been progress in this area, there is still a need for multimodal approaches that can handle both text and image inputs. To address this gap, this paper presents a multimodal veracity prediction system for automated fact-checking which was developed as part of the Factify 2 competition organized by De-Factify@AAAI 2023.
The remainder of the paper is structured as follows: Section 2 presents a brief overview of related work and section 3 describes our general framework and model architecture. Section 4 discusses the dataset supplied by the Factify 2 competition followed by an overview of our experiments in section 5. Section 6 and 7 present the final results and our conclusions, respectively.
As an essential part of automated fact verification, efective techniques for modeling
claimevidence for veracity prediction have been a hot topic and key research questions in existing
fact-checking methods. Most of the recent work focuses on using textual evidence in veracity
prediction of which there are mainly two lines of work. One direction [
Another line of work widely adopted and one of the key tasks in FEVER [
Automated multi- or cross-modal fact checking is an underdeveloped field compared to
text-based techniques. Recent developments have shown that cross-modal pre-trained models
(e.g.,VideoBERT [
We frame the Factify 2 problem as a multimodal entailment task as in the previous submission
[
Our system architecture follows a standard two-stage claim verification approach as established through various shared tasks in recent years, typically FEVER[34], FEVER 2.0 [35], FEVEROUS [36] and SCIVER [37]. First, a textual evidence retrieval component identifies from a given document the evidence passages most relevant to the corresponding claim text. Then, a transformer based cross-modal model is trained on all the input across multimodalities including selected evidence passages text, claim text, claim image, document image, claim OCR text and document OCR text to predict five multimodal entailment categories with respect to the multimodal claim. A pre-trained cross-modal model (i.e. CLIP) and a pre-trained text embedding model are both employed in the embedding layer in order to learn a cross-modal matching model using both unified-multimodal and unimodal representations. Overall, the implemented architecture adopts a list-wise concatenation strategy [38] which is one of common strategies in most recent sequence-to-sequence SoTA veracity prediction models.
In evidence retrieval, ‘multi-qa-mpnet-base-dot-v1‘ 3 and is employed to compute embeddings for both claim text and document text at passage level. In terms of passage granularity, both paragraph- and sentence-level retrieval have been experimented with (see Section 5). This is a SBERT model based on the MPNet architecture [39] and is trained on a Question-Answer (QA) dataset with 215M QA pairs from diverse sources. The model was tuned for a semantic search using a dot-product score function in order to find relevant passages corresponding to a given query. The model encodes text into a 768-d vector and supports 512 maximum number of tokens.
Regarding the similarity computation and semantic search, we use a simple dot product with the normalised SBERT embeddings (as proxy to cosine similarity) which enables a quick and eficient passage ranking and scalability of up to about 1 Million entries.
Top passages obtained from the semantic search are then re-ranked based on their relevancy to the claim text and concatenated into a longer text snippet before being fed into the cross-modal veracity prediction model.
Our embedding layer consists of a cross-modal encoder and a unimodal text encoder. We hypothesize that modeling solely on text-to-text interaction (i.e., text premise and hypothesis) can supplement the modeling solely on cross-modal premise and hypothesis interaction and vice versa. This architecture facilitates the measuring of multimodal semantic relatedness in this multimodal fact checking task by mapping more textual alignment signals into subsequent semantic space. This considers that text specific models can capture more accurate and semantically meaningful word- or sentence-level alignment.
The cross-modal encoder is implemented with a pre-trained CLIP model that aims to map visual and text embeddings into a common space. The ViT-B/32 variant (ViT-Base with patch size 32) is chosen in this work because of its smaller amount of parameters, less FLOPS and greater inference speed. ViT-B/32 consists of a text encoder and an image encoder which are used to encode text inputs (including claim text, evidentiary passage and two images OCR text) and image inputs (including claim image and document image) respectively before concatenating into a 6 × 512 matrix as a single input to the subsequent transformer encoder. The CLIP architecture allows for a maximum input text length of 77 tokens. The pre-trained Word2vec model ("Word2vec Google News 300") [40] is adopted as a unimodal text encoder. It encodes the concatenated text sequence of claim and document evidentiary passage text, and obtains a 300-D feature vector for each token. Zero-padding is applied to match the longest sentence in the training set. Both the pre-trained CLIP and Word2Vec embedding model were not fine-tuned.
The second component of veracity prediction is based on the well established Transformer Encoder (TE) architecture, which heavily relies on the concept of self-attention [41] to efectively 3The model is available on the Hugging Face hub, accessible via https://huggingface.co/sentencetransformers/multi-qa-mpnet-base-dot-v1 model higher-order interactions and context in an input. Recent research has shown that multihead self-attention mechanisms and transformer architectures are computationally eficient and accurate in this regard. The self-attention mechanisms of the TE encoder allows for simple but powerful reasoning that can identify hidden relationships between vector entities, regardless of whether they are visual or textual in nature. Therefore, our cross-modal veracity prediction model is implemented based on self-attention mechanisms to learn the joint distribution of text representations of claim-document text pair and cross-modal feature representations of all modalities contained in claim and document.
Specifically, the claim and document embeddings of joint input by CLIP and text input by text embedding layer are passed through two separate transformer encoders [41] consisting of identical sequential blocks of a multi-head attention (MHA) and a fully connected feedforward network (FFN). Within each transformer encoder, multiple blocks allows for a deeper understanding of the inputs. For each block the input is passed through a multi-head attention layer of which the output is added to the initial input such that information in the initial sequence is not lost. Layer normalization is applied to the output to allow for faster training and small regularization i.e. = LayerNorm( + MHA()). The output is then passed to a feed-forward network to allow for more model complexity. The output is again added to the original input and layer normalization is applied i.e. = LayerNorm( + FFN()). The output of the final block (i.e., the output of each transformer encoder in the diagram) is passed through an adaptive max pooling layer to reduce the output dimensions. The output of two separate transformer encoders are then concatenated before feeding into a MLP classifier for the five category prediction. The five categories probabilities are obtained from the final output softmax layer.
The Factify 2 dataset created and supplied by the organisers covers a train, validation, and test set. The train set contains 35000 data pairs, while the validation and test sets each contain 7500 data pairs. Each data pair consists of a claim and a document, each of which comprises an image, a text, and an OCR text extracted from the image. The data pairs are annotated with one label from 5 categories including Support_Multimodal, Support_Text, Refute, Insuficient_Multimodal, or Insuficient_Text.
The training set text and OCR text length distributions are represented in Figures 2 and 3. The text length distribution varies between the claim and document text, with the document text that tends to be much longer. This is expected as it is used to verify the claim. From Figure 2 (a), we can can see that claim text is much shorter and less varied for the Refute category than for the rest of the categories, which all have similar claim text length distributions. Figure 2 (b) shows that the Support_Multimodal and Support_Text categories have the larger spread of document text lengths and also the longest document text lengths. The two Insuficient categories have on average a smaller document text length, and Refute has the smallest variance and maximum length in document text length.
Considering the claim OCR length we see from Figure 3 that the Refute category has a much larger claim OCR length distribution and maximum length than any other category. The second largest claim OCR length distributions are the Support_Text and the Insuficient_Text categories, which then leaves the two Multimodal categories with the shortest claim OCR text lengths. The document OCR length distribution is very similar to that of the claim OCR, from Figure 3b we see the only real diference is that the two Text categories have a smaller document OCR length distrubution than that of the claim OCR.
(a) Claim Text Length (b) Document Text Length
An image similarity investigation was conducted in order to gain an intuition of the similarity
between the claim and document images for each category. Using image pairwise CLIP
embeddings we calculate a similarity score and analyse it per category. Figures 4a and 4b illustrate that
the similarity between the claim and document image is comparatively higher within the
categories for Support_Multimodal and Insuficient_Multimodal than the other categories. The label
correlation with similarity of image pairs has largely increased compared to factity 1 dataset
[
The multimodal CLIP similarity among multimodal claim and doc pairs is explored to investigate our hypothesis that doc image should contain content that is related to the claim in order to entail either support or refute verdict decisions. Figures 5a and 5b depict the cosine similarity scores between the claim text and document image. It is noticeable that “Support_Multimodal” presents the highest pairwise similarity correlation between label and claim-evidence pair. “Insuficient text” have the lowest pairwise similarity correlation, although our initial hypothesis was that “Insuficient_Multimodal” should have the lowest value. This analysis suggests that diferentiating between the diferent categories based on the claim text and document image correlation could be challenging.
In terms of correlation between the claim image and document text, due to the maximum text sequence constraints with CLIP, text access maximum length is truncated. Consequently, longer context of document text is not incorporated in this analysis. As shown in Figure 6a and 6c, there is low degree of similarity correlation across the five categories, among which the "Refute" category shows highest similarity correlation.
Lastly, Figure 6b and Figure 6d show the similarity correlation between the claim image and the claim text, and show no significant deviation in similarity scores of diferent categories when the claim image and claim text are compared to each other. For the purpose of this task and this dataset, we hypothesize that the claim image should provide supplementary information to the claim text.
(a) Claim Image and Document Image Similarity Score Histogram
(b) Claim Image and Document Image Similarity Boxplot
To validate and optimize the efect of evidence retrieval, we attempt to experiment with our model with 1) including or excluding evidence selection; 2) varying the length of evidence doc text sorted by evidence retriever; 3) passage ranking at paragraph level versus sentence level; 4) text-to-text alignment with SBERT versus cross-modal alignment with CLIP. Both SBERT and CLIP is used to rank evidence doc with paragraph and sentence level; 5) if SBERT model trained on QA dataset perform better than general purpose SBERT model. Note that ranking (a) Claim Text and Document Image Similarity Score Histogram
(b) Claim Text and Document Image Similarity Boxplot at paragraph level on top <5 or sentence level on top <5 is only an option for CLIP due to its maximum allowed length restriction.
For two transformer encoders, we choose an empirical setting of four heads in two MHAs. The number of sequential MHA and feed-forward network blocks per embedding input is = 2. All our experiments are trained on a 3-layered MLP and the number of nodes per layer are set to 3072, 1024 and 5, respectively. A dropout of 0.5 and ReLU activations are applied between the MLP layers.
Preliminary experiments conducted in this work are elaborated in details as follows: • "model_w/o_ER": to validate the efectiveness with evidence retrieval, we remove evidence (a) Claim Image and Document Text Similarity Score His- (b) Claim Image and Claim Text Similarity Score Histogram togram (c) Claim Image and Document Text Similarity Score
Boxplot
(d) Claim Image and Claim Text Similarity Score
Boxplot
retrieval in our system and provide original document text to "Cross-modal veracity
prediction model".
• "SBERT_sentence_ER_top5": One of the "top" 4 performing general purpose SBERT model
("all-MiniLM-L6-v2") is chosen in our experiment. This is an all-round model tuned for
many use cases and 5 times faster while ofering good quality compared to the best
all-round model "all-mpnet-base-v2". The model is trained on a large and diverse dataset
of over 1 billion training pairs and also fine-tuned for dot-product score function suitable
for cosine similarity. The use of the all-round model allows us to evaluate the value of
adopting QA fine-tuned counterpart that we hypothesize to be the optimal solution. Top
4The best performing general purpose model is selected with a sorted list of model performances and
recommended use cases provided by SBERT, accessible via https://www.sbert.net/docs/pretrained_models.html
5 sorted sentences sorted by all-round SBERT model is configured in this setting.
• "SBERT_sentence_ER_top10": Top 10 sorted sentence sorted by all-round SBERT model
is configured in this setting.
• "SBERT_sentence_ER_top15": Top 15 sorted sentence sorted by all-round SBERT model
is configured in this setting.
• "SBERT-QA_paragraph_ER_top5": SBERT QA dataset fine-tuned model (as described
in 3.3) is adopted in this setting to obtain top 5 paragraphs as evidentiary passages for
veracity inference in this setting.
• "SBERT-QA_sentence_ER_top5": Top 5 sentences sorted by SBERT QA model and selected
as evidentiary passages in this setting.
• "BigBird_w/o_ER": To evaluate the value of evidence selection against the long context
modeling solution, the Google’s BigBird pre-trained model fine-tuned on Factity dataset
from last year [
For our experiment, the model was trained up to 80 epochs with early stopping on minimum validation loss by minimizing the cross-entropy loss function using the adaptive AdamW optimizer [42] with an initial learning rate of = 1e− 4 and epsilon = 1e− 8 with batch size ℎ = 16. Early stopping patience is set to 5. A linear decreasing learning rate scheduler was used including = 438 warming up training steps during which the learning rate increased linearly to the chosen learning rate.
We have found that data scraping errors lead to invalid doc text content in the development dateset provided by organiser with 463 and 114 invalid samples in train and val set respectively. There also are 112 invalid samples in test set. This results in document text containing only "We’ve detected that JavaScript is disabled in this browser ...". The invalid samples are removed from our training data.
The best model results in preliminary experiments described in section 5 are presented in Table 1, Table 2 and Table 3 respectively.
Firstly, the Table 1 shows that our veracity model without ER exhibits a reasonably good performance and utilising the long sequence model (BigBird) for text embeddings improves the base model with a small margin, by 1% for all categories except "Refute". As comparison, further experiments with ER are conducted of which the results are presented in Table 2 and Table 3. The results in Table 2 indicate that all-round SBERT based evidence selection does not provide obvious performance improvement based on current preliminary explorations covering three top K sentences settings (K=5, 10, 15). In contrast, SERT-QA based model achieves big marginal improvement at both paragraph and sentence level. Our experiments covers both top 5 paragraphs and sentences, which improves best base model (without ER) by 5-way Classification Results of experiments without ER on val set model_w/o_ER
BigBird_w/o_ER 1% and 2% respectively. Final results across 7 diferent experiment setup shows that combining SBERT-QA at top K sentence-level evidence passage retrieval achieves optimal performance compared to the base model without ER and the use of all-round SBERT model. The best model "SBERT-QA_sentence_ER_top5" obtains 0.79 weighted average F1 at the 20th epoch. R 5-way Classification Results of experiments with all-round SBERT + ER on val set
SBERT_sentence_ER_top5 SBERT_sentence_ER_top10 SBERT_sentence_ER_top15
P R F1 P R F1 P R F1 5-way Classification Results of experiments with SBERT-QA + ER on val set
The final test set results and competition leaderboard are presented in Table 4. The results show that top 3 participating systems achieves similar performance and our system is ranked at 3rd place with a small margin (by 0.028) to the top performing system. Please refer to [43] for the competition details.
In this research, we present our multimodal fact checking system that is submitted to the
DeFactify 2023 competition. The system consists of various components, including a multimodal
fact checking dataset, a QA-enhanced evidence passage retrieval component, and a
Transformerbased cross-modal sequence-to-sequence veracity prediction model. Our findings from the
De-Factify 2023 competition show that recent advances in pre-trained cross-modal models, such
as CLIP, have strong zero-shot or few-shot capabilities and can be efectively transferred to
a variety of downstream tasks, including multimodal fact checking. However, there is still a
need for more efective techniques for multimodal modeling and explainability, particularly
in regards to learning finer-grained cross-modal representations by jointly modeling
intraand inter-modality relationships and aligning vision regions with sentence words or entities.
Additionally, more focus should be placed on real-world challenges that involve handling large
amounts of textual and multimodal information from multiple sources and domains for claim
verification. There is also a need for techniques that can efectively handle more complex
and nuanced real-world scenarios, such as those involving sarcasm, irony, and misleading
context. The dificulties in creating large and high-quality multimodal fact checking datasets
that accurately reflect real-world scenarios (e.g., insuficient/leaked evidence), as identified in
previous work [
org/abs/1711.05101. doi:10.48550/ARXIV.1711.05101. [43] S. Suryavardan, S. Mishra, M. Chakraborty, P. Patwa, A. Rani, A. Chadha, A. Reganti, A. Das, A. Sheth, M. Chinnakotla, A. Ekbal, S. Kumar, Findings of factify 2: multimodal fake news detection, in: proceedings of defactify 2: second workshop on Multimodal Fact-Checking and Hate Speech Detection, CEUR, 2023. [44] M. Glockner, Y. Hou, I. Gurevych, Missing counter-evidence renders nlp fact-checking unrealistic for misinformation, arXiv preprint arXiv:2210.13865 (2022).