The Two Sides of the Coin: Hallucination Generation and Detection with LLMs as Evaluators for LLMs

Hallucination detection in Large Language Models (LLMs) is crucial for ensuring their reliability. This work presents our participation in the CLEF ELOQUENT HalluciGen shared task, where the goal is to develop evaluators for both generating and detecting hallucinated content. We explored the capabilities of four LLMs: Llama 3, Gemma, GPT-3.5 Turbo, and GPT-4, for this purpose. We also employed ensemble majority voting to incorporate all four models for the detection task. The results provide valuable insights into the strengths and weaknesses of these LLMs in handling hallucination generation and detection tasks.

Evaluation of Temporal Change in IR Test Collections

Information retrieval systems have been evaluated using the Cranfield paradigm for many years. This paradigm allows a systematic, fair, and reproducible evaluation of different retrieval methods in fixed experimental environments. However, real-world retrieval systems must cope with dynamic environments and temporal changes that affect the document collection, topical trends, and the individual user's perception of what is considered relevant. Yet, the temporal dimension in IR evaluations is still understudied. To this end, this work investigates how the temporal generalizability of effectiveness evaluations can be assessed. As a conceptual model, we generalize Cranfield-type experiments to the temporal context by classifying the change in the essential components according to the create, update, and delete operations of persistent storage known from CRUD. From the different types of change different evaluation scenarios are derived and it is outlined what they imply. Based on these scenarios, renowned state-of-the-art retrieval systems are tested and it is investigated how the retrieval effectiveness changes on different levels of granularity. We show that the proposed measures can be well adapted to describe the changes in the retrieval results. The experiments conducted confirm that the retrieval effectiveness strongly depends on the evaluation scenario investigated. We find that not only the average retrieval performance of single systems but also the relative system performance are strongly affected by the components that change and to what extent these components changed.

Context-Driven Interactive Query Simulations Based on Generative Large Language Models

Simulating user interactions enables a more user-oriented evaluation of information retrieval (IR) systems. While user simulations are cost-efficient and reproducible, many approaches often lack fidelity regarding real user behavior. Most notably, current user models neglect the user's context, which is the primary driver of perceived relevance and the interactions with the search results. To this end, this work introduces the simulation of context-driven query reformulations. The proposed query generation methods build upon recent Large Language Model (LLM) approaches and consider the user's context throughout the simulation of a search session. Compared to simple context-free query generation approaches, these methods show better effectiveness and allow the simulation of more efficient IR sessions. Similarly, our evaluations consider more interaction context than current session-based measures and reveal interesting complementary insights in addition to the established evaluation protocols. We conclude with directions for future work and provide an entirely open experimental setup.

Simulating Users in Interactive Web Table Retrieval

Considering the multimodal signals of search items is beneficial for retrieval effectiveness. Especially in web table retrieval (WTR) experiments, accounting for multimodal properties of tables boosts effectiveness. However, it still remains an open question how the single modalities affect user experience in particular. Previous work analyzed WTR performance in ad-hoc retrieval benchmarks, which neglects interactive search behavior and limits the conclusion about the implications for real-world user environments. To this end, this work presents an in-depth evaluation of simulated interactive WTR search sessions as a more cost-efficient and reproducible alternative to real user studies. As a first of its kind, we introduce interactive query reformulation strategies based on Doc2Query, incorporating cognitive states of simulated user knowledge. Our evaluations include two perspectives on user effectiveness by considering different cost paradigms, namely query-wise and time-oriented measures of effort. Our multi-perspective evaluation scheme reveals new insights about query strategies, the impact of modalities, and different user types in simulated WTR search sessions.

Validating Synthetic Usage Data in Living Lab Environments

Evaluating retrieval performance without editorial relevance judgments is challenging, but instead, user interactions can be used as relevance signals. Living labs offer a way for small-scale platforms to validate information retrieval systems with real users. If enough user interaction data are available, click models can be parameterized from historical sessions to evaluate systems before exposing users to experimental rankings. However, interaction data are sparse in living labs, and little is studied about how click models can be validated for reliable user simulations when click data are available in moderate amounts. This work introduces an evaluation approach for validating synthetic usage data generated by click models in data-sparse human-in-the-loop environments like living labs. We ground our methodology on the click model's estimates about a system ranking compared to a reference ranking for which the relative performance is known. Our experiments compare different click models and their reliability and robustness as more session log data becomes available. In our setup, simple click models can reliably determine the relative system performance with already 20 logged sessions for 50 queries. In contrast, more complex click models require more session data for reliable estimates, but they are a better choice in simulated interleaving experiments when enough session data are available. While it is easier for click models to distinguish between more diverse systems, it is harder to reproduce the system ranking based on the same retrieval algorithm with different interpolation weights. Our setup is entirely open, and we share the code to reproduce the experiments.

Evaluating Temporal Persistence Using Replicability Measures

In real-world Information Retrieval (IR) experiments, the Evaluation Environment (EE) is exposed to constant change. Documents are added, removed, or updated, and the information need and the search behavior of users is evolving. Simultaneously, IR systems are expected to retain a consistent quality. The LongEval Lab seeks to investigate the longitudinal persistence of IR systems, and in this work, we describe our participation. We submitted runs of five advanced retrieval systems, namely a Reciprocal Rank Fusion (RRF) approach, ColBERT, monoT5, Doc2Query, and E5, to both sub-tasks. Further, we cast the longitudinal evaluation as a replicability study to better understand the temporal change observed. As a result, we quantify the persistence of the submitted runs and see great potential in this evaluation method.

Online Information Retrieval Evaluation using the STELLA Framework

Involving users in early phases of software development has become a common strategy as it enables developers to consider user needs from the beginning. Once a system is in production, new opportunities to observe, evaluate and learn from users emerge as more information becomes available. Gathering information from users to continuously evaluate their behavior is a common practice for commercial software, while the Cranfield paradigm remains the preferred option for Information Retrieval (IR) and recommendation systems in the academic world. Here we introduce the Infrastructures for Living Labs STELLA project which aims to create an evaluation infrastructure allowing experimental systems to run along production web-based academic search systems with real users. STELLA combines user interactions and log files analyses to enable large-scale A/B experiments for academic search.

ir_metadata: An Extensible Metadata Schema for IR Experiments

The information retrieval (IR) community has a strong tradition of making the computational artifacts and resources available for future reuse, allowing the validation of experimental results. Besides the actual test collections, the underlying run files are often hosted in data archives as part of conferences like TREC, CLEF, or NTCIR. Unfortunately, the run data itself does not provide much information about the underlying experiment. For instance, the single run file is not of much use without the context of the shared task's website or the run data archive. In other domains, like the social sciences, it is good practice to annotate research data with metadata. In this work, we introduce ir_metadata - an extensible metadata schema for TREC run files based on the PRIMAD model. We propose to align the metadata annotations to PRIMAD, which considers components of computational experiments that can affect reproducibility. Furthermore, we outline important components and information that should be reported in the metadata and give evidence from the literature. To demonstrate the usefulness of these metadata annotations, we implement new features in repro_eval that support the outlined metadata schema for the use case of reproducibility studies. Additionally, we curate a dataset with run files derived from experiments with different instantiations of PRIMAD components and annotate these with the corresponding metadata. In the experiments, we cover reproducibility experiments that are identified by the metadata and classified by PRIMAD. With this work, we enable IR researchers to annotate TREC run files and improve the reuse value of experimental artifacts even further.

Overview of LiLAS 2021 -- Living Labs for Academic Search

The Living Labs for Academic Search (LiLAS) lab aims to strengthen the concept of user-centric living labs for academic search. The methodological gap between real-world and lab-based evaluation should be bridged by allowing lab participants to evaluate their retrieval approaches in two real-world academic search systems from life sciences and social sciences. This overview paper outlines the two academic search systems LIVIVO and GESIS Search, and their corresponding tasks within LiLAS, which are ad-hoc retrieval and dataset recommendation. The lab is based on a new evaluation infrastructure named STELLA that allows participants to submit results corresponding to their experimental systems in the form of pre-computed runs and Docker containers that can be integrated into production systems and generate experimental results in real-time. Both submission types are interleaved with the results provided by the productive systems allowing for a seamless presentation and evaluation. The evaluation of results and a meta-analysis of the different tasks and submission types complement this overview.

Evaluating Elements of Web-based Data Enrichment for Pseudo-Relevance Feedback Retrieval

In this work, we analyze a pseudo-relevance retrieval method based on the results of web search engines. By enriching topics with text data from web search engine result pages and linked contents, we train topic-specific and cost-efficient classifiers that can be used to search test collections for relevant documents. Building upon attempts initially made at TREC Common Core 2018 by Grossman and Cormack, we address questions of system performance over time considering different search engines, queries, and test collections. Our experimental results show how and to which extent the considered components affect the retrieval performance. Overall, the analyzed method is robust in terms of average retrieval performance and a promising way to use web content for the data enrichment of relevance feedback methods.