Generative AI (GenAI) and Natural Language Processing (NLP) have advanced significantly in recent years, exhibiting breakthroughs and pushing the bar of accuracy rates in text mining. Cascade efects have been observed in many application domains, spanning text analysis, question answering, classification, and new textual content generation. The latter has allowed many end-users to perceive AI as ready-to-go solutions to optimise their daily workflow. However, dark and bright sides lurk behind textual content generation, as trustworthy and unverified content can be efortlessly generated. That has fuelled a significant challenge in our society: fake news. Although fake news has existed for a while, it remains an unsolved issue. Generative AI has brought it to a new level by enabling the automated production of large volumes of high-quality, individually targeted fake content. Our work is part of the HeReFaNMi (Health-Related Fake News Mitigation) project, which focuses on health-related fake news mitigation by using NLP, Language Models, and a Retrieval-Augmented Generation (RAG) system. We propose a new chunking mechanism that streamlines the overall RAG framework pipeline. BERT and BERT+RAG have been compared on the health-related fake news classification task on a dataset of 2000 health-related articles equally split into two categories ('fake' and 'credible'). Preliminary experimental results reveal improvements in AIxPAC: Workshop on Artificial Intelligence for Perception and Artificial Consciousness, November 25-28, 2024, Bolzano, Italy ∗Corresponding author. †These authors contributed equally.
The advent of language models has changed the field of NLP and artificial intelligence making a
meaningful contribution to diverse domains, going from creating human-like text to powering chatbots. These
models have proven an impressive capability to handle dificult tasks like translation, summarization,
and even creative writing [
Language Models [
CEUR
ceur-ws.org
attention principles. These represent a breakthrough in generative AI, allowing predicting the next
word in a sequence [
The ability of Language Models to generate human-like text makes them a powerful tool, but it
also makes them susceptible to misuse [
This paper presents a new contribution through the integration of a new chunking mechanism into the RAG framework [16]. The chunking system allows for the processing of longer and more complex text by breaking them into smaller and interpretable chunks that fit the context window size [ 17, 18], and sending them to the Language Models as context. Our approach also leverages the power of prompt engineering to fine-tune the model’s inputs for more accurate.
Our contribution represents a meaningful efort to adapt and improve RAG systems to combat healthcare-related fake news. Through a retrieval-based system, we aim to provide a robust solution to the growing problem. This paper will explore the technical aspects of our system and the potential of our approach to mitigate these problems efectively through an improved novel chunking strategy.
When building applications that rely on natural language processing (NLP) tasks, such as semantic search or document summarization, one of the most critical aspects is ensuring that the text is represented in a way that maintains its meaning and relevance. Chunking is an essential technique in this regard, particularly when working with Language Models that have token limits. Previous work has explored various chunking techniques to handle these limitations efectively [ 19, 20].
Fixed-size Chunking is the most straightforward approach, where the document is split into equally sized segments based on a predefined number of tokens or characters. Fixed-size chunking is easy to implement and works well in many scenarios [20]. However, this approach risks losing semantic context if the division happens mid-sentence or mid-idea, potentially reducing the efectiveness of Language Models in downstream tasks like retrieval and summarization [17].
Content-aware strategies focus on partitioning a document based on its inherent structure, such as sentences, paragraphs, or sections, allowing the system to preserve meaningful boundaries and enhance both retrieval and processing performance.
Sentence Splitting: This method ensures that each chunk consists of complete sentences, thus maintaining readability and coherence. Sentence splitting is especially advantageous in tasks like summarization and question answering, where preserving sentence-level integrity is crucial for maintaining context and meaning [21].
Recursive Chunking: This technique involves dividing documents using predefined separators (e.g., paragraph breaks, sentence boundaries) while ensuring that semantically relevant content is preserved. Recursive chunking is flexible, with parameters such as chunkSize (defining the maximum allowable chunk size) and chunkOverlap (controlling the degree of content shared between adjacent chunks). This approach has been particularly efective for processing long documents, ensuring that semantic continuity is maintained within and across chunks, thereby improving document understanding and task performance [22].
Documents written in markup languages (e.g., Markdown, HTML, LaTeX) implement tagging systems that categorize and structure text into both semantic and syntactic units. This intrinsic formatting ofers a unique advantage for advanced chunking techniques [ 23], specifically designed to handle semi-structured data while accommodating its flexible schema. By harnessing the structured metadata and hierarchical markers, these techniques ensure that meaningful chunk boundaries are preserved, thereby improving the eficiency and accuracy of downstream processes such as information retrieval, content summarization, and data extraction. This structured approach enhances the interpretability of the data and ensures a more coherent representation of complex document structures [24].
In this work, we present a new approach to enhance the performance of the Retrieval-Augmented Generation (RAG) by changing how the articles are split and retrieved. Our proposed method encompasses the following key steps (see Figure 1): • Sentence Splitting: the article is divided into paragraphs. This step allows a finer-grained retrieval process by which the model can work on smaller, more focused parts of information. • Cosine Similarity Calculation: We calculate the cosine similarity between the sentence embeddings to find semantically similar sentences. Those with high cosine similarity scores are grouped together to form coherent content chunks. The cosine similarity between two sentence embeddings, and , is calculated as: (1) (2) cosine_similarity( , ) = ⋅ ‖ ‖‖ ‖ • Chunk Creation: Sentences with high similarity scores are merged into chunks, in other words, similar ideas are grouped into one chunk. These chunks become the primary retrieval units, containing semantically related content, which improves the precision of the retrieved context. • Re-Embedding of Chunks: After the chunks are created, each chunk is re-embedded using a semantic model. This process captures a more meaningful representation of the chunk in the latent space, This chunk-based approach improves the retrieval process by making sure that the retrieved content is more focused and semantically coherent.
The chunk-based approach enhances retrieval by ensuring the retrieved content is more focused and semantically coherent. A more contextually relevant document can then benefit the generated output. The following subsections describe how these modifications impact the retrieval and generation components of the RAG model.
The retriever, (|) , now operates over these newly created chunks. We maintain the Dense Passage Retriever (DPR) bi-encoder architecture. Still, instead of retrieving entire documents, the model retrieves top- chunks based on the cosine similarity between the input query and each chunk embedding . The retrieval probability is formalized as: (|) ∝
exp(cosine_similarity(, )) where cosine_similarity(, ) measures the semantic similarity between the query and the chunk . This ensures that only the most relevant chunks are selected for generation. Enhanced RAG Output
Improved Retrieval Process
Re-Embedding of
Chunks
We employ two variants of the generation model: RAG-Sequence and RAG-Token [
A single retrieved chunk generates the entire output sequence in the RAG-Sequence Model. The probability of generating the sequence is marginalized over the top- retrieved chunks: RAG-Sequence( |) ≈ (|) ( |, ) The generator produces each token based on the retrieved chunk and the previous tokens: ( |, ) = 1∶−1 )
Then, we implement RAG-Token to allow the model to select a diferent chunk for each token. The probability of generating the output sequence is defined as: RAG-Token( |) ≈ ∑ (|) ( |, , 1∶−1 )
The above-described approach is dynamic and allows the model to select the most relevant chunk for generating each token in the sequence.
By using cosine similarity to group similar sentences into chunks and re-embedding these chunks, we significantly improve retrieval precision. Each chunk contains semantically consistent information, making the retrieval process more focused and reducing irrelevant or noisy content. These modifications benefit both RAG-Sequence and RAG-Token models, leading to enhanced performance in sequence ∑ ∈ top-K((⋅|)) ∏ (
|, , =1 ∏ =1 ∈ top-K((⋅|)) (3) (4) (5) generation and classification tasks.
In this section, we present the results of our experiments comparing the performance of BERT[25] with and without the use of Retrieval-Augmented Generation (RAG). We employed the RAG-Token model in our experiments over a dataset of 2000 news articles (equally balanced onto ’fake’ and ’credible’ categories) collected as part of the HeReFaNMi project 1, funded by NGI Search. The dataset, which is publicly available, was gathered using a web scraping framework and prompt engineering techniques, focusing on the classification of credible and fake news.
We evaluated the BERT [25] base model on this dataset in two scenarios: (1) without RAG, and (2) with RAG, where the model retrieves relevant text chunks before making classification decisions. The results of these experiments are summarized in Table 1:
As seen in Table 1, BERT [25] without RAG achieves an accuracy of 65.75 percent, with a recall of 31.16 percent and an F1 score of 0.4751 percent. In contrast, when using RAG, the accuracy improves to 70.10 percent, while the recall increases significantly to 89.8 percent. This substantial improvement in both accuracy and recall when using RAG indicates that the model becomes more sensitive and efective in identifying relevant instances.
The F1 score, which balances precision and recall, is 0.4751 without RAG and increases to 0.767 with RAG. This improvement in the F1 score reflects that RAG not only enhances the model’s recall but also results in better overall performance by improving precision, thereby boosting both recall and accuracy metrics.
One of the major advantages of using BERT[25], especially without RAG, is its eficiency in terms of computational resources. BERT [25] is lightweight and requires fewer resources to train and fine-tune than larger models such as LLaMA-3[26]. LLaMA-3, while powerful, demands extensive computational resources and large datasets, which are not always available or feasible to use. In contrast, BERT [25] can perform well even with smaller datasets and fewer computational requirements, making it a suitable choice for resource-constrained environments.
However, incorporating RAG into BERT [25] adds complexity to the model. Although RAG boosts recall, it introduces additional retrieval steps, which increase computational overhead. Despite this, RAG’s ability to provide relevant context to the model helps enhance its performance, especially when ifne-tuning the model on tasks requiring retrieval of external information.
Our experiments show that BERT [25], when combined with RAG, becomes much better at recalling relevant information but at the cost of precision. This suggests that RAG is beneficial for cases where recall is critical (e.g., ensuring that all potential relevant information is retrieved), but further fine-tuning or balancing mechanisms are required to maintain high accuracy.
Future work could focus on optimizing the integration of RAG with BERT [25] to reduce the trade-of between recall and accuracy, perhaps by fine-tuning the retrieval process or introducing filtering mechanisms to prevent irrelevant information from being included in the classification process. Overall, while BERT [25] with RAG shows promise, particularly for recall-focused tasks, improvements can be made to enhance its overall performance. 1Health-care Related Fake NEWS Mitigation Project (HeReFaNMi)
In conclusion, the integration of Retrieval-Augmented Generation (RAG) into models like BERT [25] presents a promising approach for improving the detection of health-related fake news. By leveraging the enhanced retrieval mechanism provided by RAG, the system can access more accurate and contextually relevant external information, leading to better recall in identifying misinformation. Our results demonstrate the trade-of between recall and accuracy, with RAG significantly boosting recall at the cost of introducing more false positives. This underscores the importance of refining retrieval processes and balancing model precision and recall for practical applications. Future work could focus on optimizing this balance and further improving the resource eficiency of such models. Ultimately, our proposed method holds significant potential for combatting misinformation, particularly in critical areas such as public health, where accuracy is paramount.
The contribution is funded by the grant awarded for HeReFaNMi - Health-Related Fake News Mitigation project, selected in the NGI Search.