In this paper, we investigate the problem of quantifying fairness in Retrieval-Augmented Generation (RAG) systems, particularly for complex cognitive tasks that go beyond factual question-answering. While RAG systems have demonstrated effectiveness in information extraction tasks, their fairness implications for cognitively complex tasks - including ideation, content creation, and analytical reasoning - remain under-explored. We propose a novel evaluation framework that extends IR fairness metrics by incorporating centrality-based measures to account for influence of retrieved documents on generated output beyond ranking. Our framework evaluates RAG systems across various cognitive dimensions using two ranking approaches: lexical (BM25) and dense (BGE), and language models of varying sizes. Our findings provide insights into: (1) the propagation of fairness disparities from retrieval to generation phases, and (2) the variation in system performance across different cognitive dimensions.

Large Language Models (LLMs) have the promise to revolutionize computing broadly, but their complexity and extensive training data also expose significant privacy vulnerabilities.One of the simplest privacy risks associated with LLMs is their susceptibility to membership inference attacks (MIAs), wherein an adversary aims to determine whether a specific data point was part of the model's training set.Although this is a known risk, state of the art methodologies for MIAs rely on training multiple computationally costly 'shadow models', making risk evaluation prohibitive for large models.Here we adapt a recent line of work which uses quantile regression to mount membership inference attacks; we extend this work by proposing a low-cost MIA that leverages an ensemble of small quantile regression models to determine if a document belongs to the model's training set or not.We demonstrate the effectiveness of this approach on fine-tuned LLMs of varying families (OPT, Pythia, Llama) and across multiple datasets.Across all scenarios we obtain comparable or improved accuracy compared to state of the art 'shadow model' approaches, with as little as 6% of their computation budget.We demonstrate increased effectiveness across multi-epoch trained target models, and architecture miss-specification robustness, that is, we can mount an effective attack against a model using a different tokenizer and architecture, without requiring knowledge on the target model.1

Many public pre-trained word embeddings have been shown to encode different types of biases. Embeddings are often obtained from training on large pre-existing corpora, and therefore resulting biases can be a reflection of unfair representations in the original data. Bias, in this scenario, is a challenging problem since current mitigation techniques require knowing and understanding existing biases in the embedding, which is not always possible. In this work, we propose to improve word embedding fairness by borrowing methods from the field of data privacy. The idea behind this approach is to treat bias as if it were a special type of training data leakage. This has the unique advantage of not requiring prior knowledge of potential biases in word embeddings. We investigated two types of privacy algorithms, and measured their effect on bias using four different metrics. To investigate techniques from differential privacy, we applied Gaussian perturbation to public pre-trained word embeddings. To investigate noiseless privacy, we applied vector quantization during training. Experiments show that both approaches improve fairness for commonly used embeddings, and additionally, noiseless privacy techniques reduce the size of the resulting embedding representation.

Abstract Customer reviews are an effective source of information about what people deem important in products (e.g. “strong zipper” for tents). These crowd-created descriptors not only highlight key product attributes, but can also complement seller-provided product descriptions. Motivated by this, we propose to leverage customer reviews to generate queries pertinent to target products in an e-commerce setting. While there has been work on automatic query generation, it often relied on proprietary user search data to generate query-document training pairs for learning supervised models. We take a different view and focus on leveraging reviews without training on search logs, making reproduction more viable by the public. Our method adopts an ensemble of the statistical properties of review terms and a zero-shot neural model trained on adapted external corpus to synthesize queries. Compared to competitive baselines, we show that the generated queries based on our method both better align with actual customer queries and can benefit retrieval effectiveness.

In online shopping quality is a key consideration when purchasing an item. Since customers cannot physically touch or try out an item before buying it, they must assess its quality from information gathered online. In a typical eCommerce setting, the customer is presented with seller-generated content from the product catalog, such as an image of the product, a textual description, and lists or comparisons of attributes. In addition to catalog attributes, customers often have access to customer-generated content such as reviews and product questions and answers. In a crowdsourced study, we asked crowd workers to compare product pairs from kitchen, electronics, home, beauty and office categories. In a side-by-side comparison, we asked them to choose the product that is higher quality, and further to identify the attributes that contributed to their judgment, where the attributes were both seller-generated and customer-generated. We find that customers tend to perceive more expensive items as higher quality but that their purchase decisions are uncorrelated with quality, suggesting that customers seek a trade-off between price and quality when making purchase decisions. Crowd workers placed a higher value on attributes derived from customer-generated content such as reviews than on catalog attributes. Among the catalog attributes, brand, item material and pack size were most often selected. Finally, attributes with a low correlation with perceived quality are nonetheless useful in predicting purchases in a machine-learned system.