PRISM: A Methodology for Auditing Biases in Large Language Models

Auditing Large Language Models (LLMs) to discover their biases and preferences is an emerging challenge in creating Responsible Artificial Intelligence (AI). While various methods have been proposed to elicit the preferences of such models, countermeasures have been taken by LLM trainers, such that LLMs hide, obfuscate or point blank refuse to disclosure their positions on certain subjects. This paper presents PRISM, a flexible, inquiry-based methodology for auditing LLMs - that seeks to illicit such positions indirectly through task-based inquiry prompting rather than direct inquiry of said preferences. To demonstrate the utility of the methodology, we applied PRISM on the Political Compass Test, where we assessed the political leanings of twenty-one LLMs from seven providers. We show LLMs, by default, espouse positions that are economically left and socially liberal (consistent with prior work). We also show the space of positions that these models are willing to espouse - where some models are more constrained and less compliant than others - while others are more neutral and objective. In sum, PRISM can more reliably probe and audit LLMs to understand their preferences, biases and constraints.

Evaluation of Attribution Bias in Retrieval-Augmented Large Language Models

Attributing answers to source documents is an approach used to enhance the verifiability of a model's output in retrieval augmented generation (RAG). Prior work has mainly focused on improving and evaluating the attribution quality of large language models (LLMs) in RAG, but this may come at the expense of inducing biases in the attribution of answers. We define and examine two aspects in the evaluation of LLMs in RAG pipelines, namely attribution sensitivity and bias with respect to authorship information. We explicitly inform an LLM about the authors of source documents, instruct it to attribute its answers, and analyze (i) how sensitive the LLM's output is to the author of source documents, and (ii) whether the LLM exhibits a bias towards human-written or AI-generated source documents. We design an experimental setup in which we use counterfactual evaluation to study three LLMs in terms of their attribution sensitivity and bias in RAG pipelines. Our results show that adding authorship information to source documents can significantly change the attribution quality of LLMs by 3% to 18%. Moreover, we show that LLMs can have an attribution bias towards explicit human authorship, which can serve as a competing hypothesis for findings of prior work that shows that LLM-generated content may be preferred over human-written contents. Our findings indicate that metadata of source documents can influence LLMs' trust, and how they attribute their answers. Furthermore, our research highlights attribution bias and sensitivity as a novel aspect of brittleness in LLMs.

Can We Use Large Language Models to Fill Relevance Judgment Holes?

Incomplete relevance judgments limit the re-usability of test collections. When new systems are compared against previous systems used to build the pool of judged documents, they often do so at a disadvantage due to the ``holes'' in test collection (i.e., pockets of un-assessed documents returned by the new system). In this paper, we take initial steps towards extending existing test collections by employing Large Language Models (LLM) to fill the holes by leveraging and grounding the method using existing human judgments. We explore this problem in the context of Conversational Search using TREC iKAT, where information needs are highly dynamic and the responses (and, the results retrieved) are much more varied (leaving bigger holes). While previous work has shown that automatic judgments from LLMs result in highly correlated rankings, we find substantially lower correlates when human plus automatic judgments are used (regardless of LLM, one/two/few shot, or fine-tuned). We further find that, depending on the LLM employed, new runs will be highly favored (or penalized), and this effect is magnified proportionally to the size of the holes. Instead, one should generate the LLM annotations on the whole document pool to achieve more consistent rankings with human-generated labels. Future work is required to prompt engineering and fine-tuning LLMs to reflect and represent the human annotations, in order to ground and align the models, such that they are more fit for purpose.

TREC iKAT 2023: A Test Collection for Evaluating Conversational and Interactive Knowledge Assistants

Conversational information seeking has evolved rapidly in the last few years with the development of Large Language Models (LLMs), providing the basis for interpreting and responding in a naturalistic manner to user requests. The extended TREC Interactive Knowledge Assistance Track (iKAT) collection aims to enable researchers to test and evaluate their Conversational Search Agents (CSA). The collection contains a set of 36 personalized dialogues over 20 different topics each coupled with a Personal Text Knowledge Base (PTKB) that defines the bespoke user personas. A total of 344 turns with approximately 26,000 passages are provided as assessments on relevance, as well as additional assessments on generated responses over four key dimensions: relevance, completeness, groundedness, and naturalness. The collection challenges CSA to efficiently navigate diverse personal contexts, elicit pertinent persona information, and employ context for relevant conversations. The integration of a PTKB and the emphasis on decisional search tasks contribute to the uniqueness of this test collection, making it an essential benchmark for advancing research in conversational and interactive knowledge assistants.

A Conceptual Framework for Conversational Search and Recommendation: Conceptualizing Agent-Human Interactions During the Conversational Search Process

The conversational search task aims to enable a user to resolve information needs via natural language dialogue with an agent. In this paper, we aim to develop a conceptual framework of the actions and intents of users and agents explaining how these actions enable the user to explore the search space and resolve their information need. We outline the different actions and intents, before discussing key decision points in the conversation where the agent needs to decide how to steer the conversational search process to a successful and/or satisfactory conclusion. Essentially, this paper provides a conceptualization of the conversational search process between an agent and user, which provides a framework and a starting point for research, development and evaluation of conversational search agents.

Accessibility in Information Retrieval

This paper introduces the concept of accessibility from the field of transportation planning and adopts it within the context of Information Retrieval (IR). An analogy is drawn between the fields, which motivates the development of document accessibility measures for IR systems. Considering the accessibility of documents within a collection given an IR System provides a different perspective on the analysis and evaluation of such systems which could be used to inform the design, tuning and management of current and future IR systems.

Measuring Bias in a Ranked List using Term-based Representations

In most recent studies, gender bias in document ranking is evaluated with the NFaiRR metric, which measures bias in a ranked list based on an aggregation over the unbiasedness scores of each ranked document. This perspective in measuring the bias of a ranked list has a key limitation: individual documents of a ranked list might be biased while the ranked list as a whole balances the groups' representations. To address this issue, we propose a novel metric called TExFAIR (term exposure-based fairness), which is based on two new extensions to a generic fairness evaluation framework, attention-weighted ranking fairness (AWRF). TExFAIR assesses fairness based on the term-based representation of groups in a ranked list: (i) an explicit definition of associating documents to groups based on probabilistic term-level associations, and (ii) a rank-biased discounting factor (RBDF) for counting non-representative documents towards the measurement of the fairness of a ranked list. We assess TExFAIR on the task of measuring gender bias in passage ranking, and study the relationship between TExFAIR and NFaiRR. Our experiments show that there is no strong correlation between TExFAIR and NFaiRR, which indicates that TExFAIR measures a different dimension of fairness than NFaiRR. With TExFAIR, we extend the AWRF framework to allow for the evaluation of fairness in settings with term-based representations of groups in documents in a ranked list.

The Influence of Presentation and Performance on User Satisfaction

The effectiveness of an IR system is gauged not just by its ability to retrieve relevant results but also by how it presents these results to users; an engaging presentation often correlates with increased user satisfaction. While existing research has delved into the link between user satisfaction, IR performance metrics, and presentation, these aspects have typically been investigated in isolation. Our research aims to bridge this gap by examining the relationship between query performance, presentation and user satisfaction. For our analysis, we conducted a between-subjects experiment comparing the effectiveness of various result card layouts for an ad-hoc news search interface. Drawing data from the TREC WaPo 2018 collection, we centered our study on four specific topics. Within each of these topics, we assessed six distinct queries with varying nDCG values. Our study involved 164 participants who were exposed to one of five distinct layouts containing result cards, such as "title'', "title+image'', or "title+image+summary''. Our findings indicate that while nDCG is a strong predictor of user satisfaction at the query level, there exists no linear relationship between the performance of the query, presentation of results and user satisfaction. However, when considering the total gain on the initial result page, we observed that presentation does play a significant role in user satisfaction (at the query level) for certain layouts with result cards such as, title+image or title+image+summary. Our results also suggest that the layout differences have complex and multifaceted impacts on satisfaction. We demonstrate the capacity to equalize user satisfaction levels between queries of varying performance by changing how results are presented. This emphasizes the necessity to harmonize both performance and presentation in IR systems, considering users' diverse preferences.

Ranking Heterogeneous Search Result Pages using the Interactive Probability Ranking Principle

The Probability Ranking Principle (PRP) ranks search results based on their expected utility derived solely from document contents, often overlooking the nuances of presentation and user interaction. However, with the evolution of Search Engine Result Pages (SERPs), now comprising a variety of result cards, the manner in which these results are presented is pivotal in influencing user engagement and satisfaction. This shift prompts the question: How does the PRP and its user-centric counterpart, the Interactive Probability Ranking Principle (iPRP), compare in the context of these heterogeneous SERPs? Our study draws a comparison between the PRP and the iPRP, revealing significant differences in their output. The iPRP, accounting for item-specific costs and interaction probabilities to determine the ``Expected Perceived Utility" (EPU), yields different result orderings compared to the PRP. We evaluate the effect of the EPU on the ordering of results by observing changes in the ranking within a heterogeneous SERP compared to the traditional ``ten blue links''. We find that changing the presentation affects the ranking of items according to the (iPRP) by up to 48\% (with respect to DCG, TBG and RBO) in ad-hoc search tasks on the TREC WaPo Collection. This work suggests that the iPRP should be employed when ranking heterogeneous SERPs to provide a user-centric ranking that adapts the ordering based on the presentation and user engagement.

TREC iKAT 2023: The Interactive Knowledge Assistance Track Overview

Conversational Information Seeking stands as a pivotal research area with significant contributions from previous works. The TREC Interactive Knowledge Assistance Track (iKAT) builds on the foundational work of the TREC Conversational Assistance Track (CAsT). However, iKAT distinctively emphasizes the creation and research of conversational search agents that adapt responses based on user's prior interactions and present context. The challenge lies in enabling Conversational Search Agents (CSA) to incorporate this personalized context to efficiency and effectively guide users through the relevant information to them. iKAT also emphasizes decisional search tasks, where users sift through data and information to weigh up options in order to reach a conclusion or perform an action. These tasks, prevalent in everyday information-seeking decisions -- be it related to travel, health, or shopping -- often revolve around a subset of high-level information operators where queries or questions about the information space include: finding options, comparing options, identifying the pros and cons of options, etc. Given the different personas and their information need (expressed through the sequence of questions), diverse conversation trajectories will arise -- because the answers to these similar queries will be very different. In this paper, we report on the first year of TREC iKAT, describing the task, topics, data collection, and evaluation framework. We further review the submissions and summarize the findings.