Pre-trained language models (PLMs), despite showing strong performance, can carry and increase biases, which can limit the development of fair NLP and IR systems. This research investigates the foundational origins of bias within PLMs, moving beyond detection to a detailed analysis of its formation and propagation across diverse architectures. Through a novel attention weight analysis, we reveal distinct attention patterns for biased versus neutral content, offering insights into the internal representations learned by PLMs. Our findings demonstrate a complex interplay between training data and model architecture, revealing that while the transformer's self-attention mechanism amplifies existing biases, the training data plays a crucial role in the initial encoding of bias within the model's representations.

Large amounts of medical visual data are produced in hospitals daily and made available continuously via publications in the scientific literature, representing the medical knowledge. However, it is not always easy to find the desired information and in clinical routine the time to fulfil an information need is often very limited. Information retrieval systems are a useful tool to provide access to these documents/images in the biomedical literature related to information needs of medical professionals. Shangri–La is a medical retrieval system that can potentially help clinicians to make decisions on difficult cases. It retrieves articles from the biomedical literature when querying a case description and attached images. The system is based on a multimodal retrieval approach with a focus on the integration of visual information connected to text. The approach includes a query–adaptive multimodal fusion criterion that analyses if visual features are suitable to be fused with text for the retrieval. Furthermore, image modality information is integrated in the retrieval step. The approach is evaluated using the ImageCLEFmed 2013 medical retrieval benchmark and can thus be compared to other approaches. Results show that the final approach outperforms the best multimodal approach submitted to ImageCLEFmed 2013.