Training on the Test Model: Contamination in Ranking Distillation

Neural approaches to ranking based on pre-trained language models are highly effective in ad-hoc search. However, the computational expense of these models can limit their application. As such, a process known as knowledge distillation is frequently applied to allow a smaller, efficient model to learn from an effective but expensive model. A key example of this is the distillation of expensive API-based commercial Large Language Models into smaller production-ready models. However, due to the opacity of training data and processes of most commercial models, one cannot ensure that a chosen test collection has not been observed previously, creating the potential for inadvertent data contamination. We, therefore, investigate the effect of a contaminated teacher model in a distillation setting. We evaluate several distillation techniques to assess the degree to which contamination occurs during distillation. By simulating a ``worst-case'' setting where the degree of contamination is known, we find that contamination occurs even when the test data represents a small fraction of the teacher's training samples. We, therefore, encourage caution when training using black-box teacher models where data provenance is ambiguous.

Few-shot Pairwise Rank Prompting: An Effective Non-Parametric Retrieval Model

A supervised ranking model, despite its advantage of being effective, usually involves complex processing - typically multiple stages of task-specific pre-training and fine-tuning. This has motivated researchers to explore simpler pipelines leveraging large language models (LLMs) that are capable of working in a zero-shot manner. However, since zero-shot inference does not make use of a training set of pairs of queries and their relevant documents, its performance is mostly worse than that of supervised models, which are trained on such example pairs. Motivated by the existing findings that training examples generally improve zero-shot performance, in our work, we explore if this also applies to ranking models. More specifically, given a query and a pair of documents, the preference prediction task is improved by augmenting examples of preferences for similar queries from a training set. Our proposed pairwise few-shot ranker demonstrates consistent improvements over the zero-shot baseline on both in-domain (TREC DL) and out-domain (BEIR subset) retrieval benchmarks. Our method also achieves a close performance to that of a supervised model without requiring any complex training pipeline.

Top-Down Partitioning for Efficient List-Wise Ranking

Large Language Models (LLMs) have significantly impacted many facets of natural language processing and information retrieval. Unlike previous encoder-based approaches, the enlarged context window of these generative models allows for ranking multiple documents at once, commonly called list-wise ranking. However, there are still limits to the number of documents that can be ranked in a single inference of the model, leading to the broad adoption of a sliding window approach to identify the k most relevant items in a ranked list. We argue that the sliding window approach is not well-suited for list-wise re-ranking because it (1) cannot be parallelized in its current form, (2) leads to redundant computational steps repeatedly re-scoring the best set of documents as it works its way up the initial ranking, and (3) prioritizes the lowest-ranked documents for scoring rather than the highest-ranked documents by taking a bottom-up approach. Motivated by these shortcomings and an initial study that shows list-wise rankers are biased towards relevant documents at the start of their context window, we propose a novel algorithm that partitions a ranking to depth k and processes documents top-down. Unlike sliding window approaches, our algorithm is inherently parallelizable due to the use of a pivot element, which can be compared to documents down to an arbitrary depth concurrently. In doing so, we reduce the number of expected inference calls by around 33% when ranking at depth 100 while matching the performance of prior approaches across multiple strong re-rankers.

Generative Relevance Feedback and Convergence of Adaptive Re-Ranking: University of Glasgow Terrier Team at TREC DL 2023

This paper describes our participation in the TREC 2023 Deep Learning Track. We submitted runs that apply generative relevance feedback from a large language model in both a zero-shot and pseudo-relevance feedback setting over two sparse retrieval approaches, namely BM25 and SPLADE. We couple this first stage with adaptive re-ranking over a BM25 corpus graph scored using a monoELECTRA cross-encoder. We investigate the efficacy of these generative approaches for different query types in first-stage retrieval. In re-ranking, we investigate operating points of adaptive re-ranking with different first stages to find the point in graph traversal where the first stage no longer has an effect on the performance of the overall retrieval pipeline. We find some performance gains from the application of generative query reformulation. However, our strongest run in terms of P@10 and nDCG@10 applied both adaptive re-ranking and generative pseudo-relevance feedback, namely uogtr_b_grf_e_gb.

"In-Context Learning" or: How I learned to stop worrying and love "Applied Information Retrieval"

With the increasing ability of large language models (LLMs), in-context learning (ICL) has evolved as a new paradigm for natural language processing (NLP), where instead of fine-tuning the parameters of an LLM specific to a downstream task with labeled examples, a small number of such examples is appended to a prompt instruction for controlling the decoder's generation process. ICL, thus, is conceptually similar to a non-parametric approach, such as $k$-NN, where the prediction for each instance essentially depends on the local topology, i.e., on a localised set of similar instances and their labels (called few-shot examples). This suggests that a test instance in ICL is analogous to a query in IR, and similar examples in ICL retrieved from a training set relate to a set of documents retrieved from a collection in IR. While standard unsupervised ranking models can be used to retrieve these few-shot examples from a training set, the effectiveness of the examples can potentially be improved by re-defining the notion of relevance specific to its utility for the downstream task, i.e., considering an example to be relevant if including it in the prompt instruction leads to a correct prediction. With this task-specific notion of relevance, it is possible to train a supervised ranking model (e.g., a bi-encoder or cross-encoder), which potentially learns to optimally select the few-shot examples. We believe that the recent advances in neural rankers can potentially find a use case for this task of optimally choosing examples for more effective downstream ICL predictions.

Exploiting Positional Bias for Query-Agnostic Generative Content in Search

In recent years, neural ranking models (NRMs) have been shown to substantially outperform their lexical counterparts in text retrieval. In traditional search pipelines, a combination of features leads to well-defined behaviour. However, as neural approaches become increasingly prevalent as the final scoring component of engines or as standalone systems, their robustness to malicious text and, more generally, semantic perturbation needs to be better understood. We posit that the transformer attention mechanism can induce exploitable defects through positional bias in search models, leading to an attack that could generalise beyond a single query or topic. We demonstrate such defects by showing that non-relevant text--such as promotional content--can be easily injected into a document without adversely affecting its position in search results. Unlike previous gradient-based attacks, we demonstrate these biases in a query-agnostic fashion. In doing so, without the knowledge of topicality, we can still reduce the negative effects of non-relevant content injection by controlling injection position. Our experiments are conducted with simulated on-topic promotional text automatically generated by prompting LLMs with topical context from target documents. We find that contextualisation of a non-relevant text further reduces negative effects whilst likely circumventing existing content filtering mechanisms. In contrast, lexical models are found to be more resilient to such content injection attacks. We then investigate a simple yet effective compensation for the weaknesses of the NRMs in search, validating our hypotheses regarding transformer bias.

Analyzing Adversarial Attacks on Sequence-to-Sequence Relevance Models

Modern sequence-to-sequence relevance models like monoT5 can effectively capture complex textual interactions between queries and documents through cross-encoding. However, the use of natural language tokens in prompts, such as Query, Document, and Relevant for monoT5, opens an attack vector for malicious documents to manipulate their relevance score through prompt injection, e.g., by adding target words such as true. Since such possibilities have not yet been considered in retrieval evaluation, we analyze the impact of query-independent prompt injection via manually constructed templates and LLM-based rewriting of documents on several existing relevance models. Our experiments on the TREC Deep Learning track show that adversarial documents can easily manipulate different sequence-to-sequence relevance models, while BM25 (as a typical lexical model) is not affected. Remarkably, the attacks also affect encoder-only relevance models (which do not rely on natural language prompt tokens), albeit to a lesser extent.