Systematic reviews are crucial for evidence-based medicine as they comprehensively analyse published research findings on specific questions. Conducting such reviews is often resource- and time-intensive, especially in the screening phase, where abstracts of publications are assessed for inclusion in a review. This study investigates the effectiveness of using zero-shot large language models~(LLMs) for automatic screening. We evaluate the effectiveness of eight different LLMs and investigate a calibration technique that uses a predefined recall threshold to determine whether a publication should be included in a systematic review. Our comprehensive evaluation using five standard test collections shows that instruction fine-tuning plays an important role in screening, that calibration renders LLMs practical for achieving a targeted recall, and that combining both with an ensemble of zero-shot models saves significant screening time compared to state-of-the-art approaches.
We describe team ielab from CSIRO and The University of Queensland's approach to the 2023 TREC Clinical Trials Track. Our approach was to use neural rankers but to utilise Large Language Models to overcome the issue of lack of training data for such rankers. Specifically, we employ ChatGPT to generate relevant patient descriptions for randomly selected clinical trials from the corpus. This synthetic dataset, combined with human-annotated training data from previous years, is used to train both dense and sparse retrievers based on PubmedBERT. Additionally, a cross-encoder re-ranker is integrated into the system. To further enhance the effectiveness of our approach, we prompting GPT-4 as a TREC annotator to provide judgments on our run files. These judgments are subsequently employed to re-rank the results. This architecture tightly integrates strong PubmedBERT-based rankers with the aid of SOTA Large Language Models, demonstrating a new approach to clinical trial retrieval.
In the field of information retrieval, Query Likelihood Models (QLMs) rank documents based on the probability of generating the query given the content of a document. Recently, advanced large language models (LLMs) have emerged as effective QLMs, showcasing promising ranking capabilities. This paper focuses on investigating the genuine zero-shot ranking effectiveness of recent LLMs, which are solely pre-trained on unstructured text data without supervised instruction fine-tuning. Our findings reveal the robust zero-shot ranking ability of such LLMs, highlighting that additional instruction fine-tuning may hinder effectiveness unless a question generation task is present in the fine-tuning dataset. Furthermore, we introduce a novel state-of-the-art ranking system that integrates LLM-based QLMs with a hybrid zero-shot retriever, demonstrating exceptional effectiveness in both zero-shot and few-shot scenarios. We make our codebase publicly available at https://github.com/ielab/llm-qlm.
Large Language Models (LLMs) demonstrate impressive effectiveness in zero-shot document ranking tasks. Pointwise, Pairwise, and Listwise prompting approaches have been proposed for LLM-based zero-shot ranking. Our study begins by thoroughly evaluating these existing approaches within a consistent experimental framework, considering factors like model size, token consumption, latency, among others. This first-of-its-kind comparative evaluation of these approaches allows us to identify the trade-offs between effectiveness and efficiency inherent in each approach. We find that while Pointwise approaches score high on efficiency, they suffer from poor effectiveness. Conversely, Pairwise approaches demonstrate superior effectiveness but incur high computational overhead. To further enhance the efficiency of LLM-based zero-shot ranking, we propose a novel Setwise prompting approach. Our approach reduces the number of LLM inferences and the amount of prompt token consumption during the ranking procedure, significantly improving the efficiency of LLM-based zero-shot ranking. We test our method using the TREC DL datasets and the BEIR zero-shot document ranking benchmark. The empirical results indicate that our approach considerably reduces computational costs while also retaining high zero-shot ranking effectiveness.
We propose the new problem of choosing which dense retrieval model to use when searching on a new collection for which no labels are available, i.e. in a zero-shot setting. Many dense retrieval models are readily available. Each model however is characterized by very differing search effectiveness -- not just on the test portion of the datasets in which the dense representations have been learned but, importantly, also across different datasets for which data was not used to learn the dense representations. This is because dense retrievers typically require training on a large amount of labeled data to achieve satisfactory search effectiveness in a specific dataset or domain. Moreover, effectiveness gains obtained by dense retrievers on datasets for which they are able to observe labels during training, do not necessarily generalise to datasets that have not been observed during training. This is however a hard problem: through empirical experimentation we show that methods inspired by recent work in unsupervised performance evaluation with the presence of domain shift in the area of computer vision and machine learning are not effective for choosing highly performing dense retrievers in our setup. The availability of reliable methods for the selection of dense retrieval models in zero-shot settings that do not require the collection of labels for evaluation would allow to streamline the widespread adoption of dense retrieval. This is therefore an important new problem we believe the information retrieval community should consider. Implementation of methods, along with raw result files and analysis scripts are made publicly available at https://www.github.com/anonymized.
Can ChatGPT provide evidence to support its answers? Does the evidence it suggests actually exist and does it really support its answer? We investigate these questions using a collection of domain-specific knowledge-based questions, specifically prompting ChatGPT to provide both an answer and supporting evidence in the form of references to external sources. We also investigate how different prompts impact answers and evidence. We find that ChatGPT provides correct or partially correct answers in about half of the cases (50.6% of the times), but its suggested references only exist 14% of the times. We further provide insights on the generated references that reveal common traits among the references that ChatGPT generates, and show how even if a reference provided by the model does exist, this reference often does not support the claims ChatGPT attributes to it. Our findings are important because (1) they are the first systematic analysis of the references created by ChatGPT in its answers; (2) they suggest that the model may leverage good quality information in producing correct answers, but is unable to attribute real evidence to support its answers. Prompts, raw result files and manual analysis are made publicly available.
Search methods based on Pretrained Language Models (PLM) have demonstrated great effectiveness gains compared to statistical and early neural ranking models. However, fine-tuning PLM-based rankers requires a great amount of annotated training data. Annotating data involves a large manual effort and thus is expensive, especially in domain specific tasks. In this paper we investigate fine-tuning PLM-based rankers under limited training data and budget. We investigate two scenarios: fine-tuning a ranker from scratch, and domain adaptation starting with a ranker already fine-tuned on general data, and continuing fine-tuning on a target dataset. We observe a great variability in effectiveness when fine-tuning on different randomly selected subsets of training data. This suggests that it is possible to achieve effectiveness gains by actively selecting a subset of the training data that has the most positive effect on the rankers. This way, it would be possible to fine-tune effective PLM rankers at a reduced annotation budget. To investigate this, we adapt existing Active Learning (AL) strategies to the task of fine-tuning PLM rankers and investigate their effectiveness, also considering annotation and computational costs. Our extensive analysis shows that AL strategies do not significantly outperform random selection of training subsets in terms of effectiveness. We further find that gains provided by AL strategies come at the expense of more assessments (thus higher annotation costs) and AL strategies underperform random selection when comparing effectiveness given a fixed annotation cost. Our results highlight that ``optimal'' subsets of training data that provide high effectiveness at low annotation cost do exist, but current mainstream AL strategies applied to PLM rankers are not capable of identifying them.
Screening prioritisation in medical systematic reviews aims to rank the set of documents retrieved by complex Boolean queries. The goal is to prioritise the most important documents so that subsequent review steps can be carried out more efficiently and effectively. The current state of the art uses the final title of the review to rank documents using BERT-based neural neural rankers. However, the final title is only formulated at the end of the review process, which makes this approach impractical as it relies on ex post facto information. At the time of screening, only a rough working title is available, with which the BERT-based ranker achieves is significantly worse than the final title. In this paper, we explore alternative sources of queries for screening prioritisation, such as the Boolean query used to retrieve the set of documents to be screened, and queries generated by instruction-based generative large language models such as ChatGPT and Alpaca. Our best approach is not only practical based on the information available at screening time, but is similar in effectiveness with the final title.
Federated online learning to rank (FOLTR) aims to preserve user privacy by not sharing their searchable data and search interactions, while guaranteeing high search effectiveness, especially in contexts where individual users have scarce training data and interactions. For this, FOLTR trains learning to rank models in an online manner -- i.e. by exploiting users' interactions with the search systems (queries, clicks), rather than labels -- and federatively -- i.e. by not aggregating interaction data in a central server for training purposes, but by training instances of a model on each user device on their own private data, and then sharing the model updates, not the data, across a set of users that have formed the federation. Existing FOLTR methods build upon advances in federated learning. While federated learning methods have been shown effective at training machine learning models in a distributed way without the need of data sharing, they can be susceptible to attacks that target either the system's security or its overall effectiveness. In this paper, we consider attacks on FOLTR systems that aim to compromise their search effectiveness. Within this scope, we experiment with and analyse data and model poisoning attack methods to showcase their impact on FOLTR search effectiveness. We also explore the effectiveness of defense methods designed to counteract attacks on FOLTR systems. We contribute an understanding of the effect of attack and defense methods for FOLTR systems, as well as identifying the key factors influencing their effectiveness.
Current methods of evaluating search strategies and automated citation screening for systematic literature reviews typically rely on counting the number of relevant and not relevant publications. This established practice, however, does not accurately reflect the reality of conducting a systematic review, because not all included publications have the same influence on the final outcome of the systematic review. More specifically, if an important publication gets excluded or included, this might significantly change the overall review outcome, while not including or excluding less influential studies may only have a limited impact. However, in terms of evaluation measures, all inclusion and exclusion decisions are treated equally and, therefore, failing to retrieve publications with little to no impact on the review outcome leads to the same decrease in recall as failing to retrieve crucial publications. We propose a new evaluation framework that takes into account the impact of the reported study on the overall systematic review outcome. We demonstrate the framework by extracting review meta-analysis data and estimating outcome effects using predictions from ranking runs on systematic reviews of interventions from CLEF TAR 2019 shared task. We further measure how closely the obtained outcomes are to the outcomes of the original review if the arbitrary rankings were used. We evaluate 74 runs using the proposed framework and compare the results with those obtained using standard IR measures. We find that accounting for the difference in review outcomes leads to a different assessment of the quality of a system than if traditional evaluation measures were used. Our analysis provides new insights into the evaluation of retrieval results in the context of systematic review automation, emphasising the importance of assessing the usefulness of each document beyond binary relevance.