Survivorship bias is the tendency to concentrate on the positive outcomes of a selection process and overlook the results that generate negative outcomes. We observe that this bias could be present in the popular MS MARCO dataset, given that annotators could not find answers to 38--45% of the queries, leading to these queries being discarded in training and evaluation processes. Although we find that some discarded queries in MS MARCO are ill-defined or otherwise unanswerable, many are valid questions that could be answered had the collection been annotated more completely (around two thirds using modern ranking techniques). This survivability problem distorts the MS MARCO collection in several ways. We find that it affects the natural distribution of queries in terms of the type of information needed. When used for evaluation, we find that the bias likely yields a significant distortion of the absolute performance scores observed. Finally, given that MS MARCO is frequently used for model training, we train models based on subsets of MS MARCO that simulates more survivorship bias. We find that models trained in this setting are up to 9.9% worse when evaluated on versions of the dataset with more complete annotations, and up to 3.5% worse at zero-shot transfer. Our findings are complementary to other recent suggestions for further annotation of MS MARCO, but with a focus on discarded queries. Code and data for reproducing the results of this paper are available in an online appendix.
Despite its troubled past, the AOL Query Log continues to be an important resource to the research community -- particularly for tasks like search personalisation. When using the query log these ranking experiments, little attention is usually paid to the document corpus. Recent work typically uses a corpus containing versions of the documents collected long after the log was produced. Given that web documents are prone to change over time, we study the differences present between a version of the corpus containing documents as they appeared in 2017 (which has been used by several recent works) and a new version we construct that includes documents close to as they appeared at the time the query log was produced (2006). We demonstrate that this new version of the corpus has a far higher coverage of documents present in the original log (93%) than the 2017 version (55%). Among the overlapping documents, the content often differs substantially. Given these differences, we re-conduct session search experiments that originally used the 2017 corpus and find that when using our corpus for training or evaluation, system performance improves. We place the results in context by introducing recent adhoc ranking baselines. We also confirm the navigational nature of the queries in the AOL corpus by showing that including the URL substantially improves performance across a variety of models. Our version of the corpus can be easily reconstructed by other researchers and is included in the ir-datasets package.
We present ir-measures, a new tool that makes it convenient to calculate a diverse set of evaluation measures used in information retrieval. Rather than implementing its own measure calculations, ir-measures provides a common interface to a handful of evaluation tools. The necessary tools are automatically invoked (potentially multiple times) to calculate all the desired metrics, simplifying the evaluation process for the user. The tool also makes it easier for researchers to use recently-proposed measures (such as those from the C/W/L framework) alongside traditional measures, potentially encouraging their adoption.
We consider the query recommendation problem in closed loop interactive learning settings like online information gathering and exploratory analytics. The problem can be naturally modelled using the Multi-Armed Bandits (MAB) framework with countably many arms. The standard MAB algorithms for countably many arms begin with selecting a random set of candidate arms and then applying standard MAB algorithms, e.g., UCB, on this candidate set downstream. We show that such a selection strategy often results in higher cumulative regret and to this end, we propose a selection strategy based on the maximum utility of the arms. We show that in tasks like online information gathering, where sequential query recommendations are employed, the sequences of queries are correlated and the number of potentially optimal queries can be reduced to a manageable size by selecting queries with maximum utility with respect to the currently executing query. Our experimental results using a recent real online literature discovery service log file demonstrate that the proposed arm selection strategy improves the cumulative regret substantially with respect to the state-of-the-art baseline algorithms. % and commonly used random selection strategy for a variety of contextual multi-armed bandit algorithms. Our data model and source code are available at ~\url{https://anonymous.4open.science/r/0e5ad6b7-ac02-4577-9212-c9d505d3dbdb/}.
Search result diversification is a beneficial approach to overcome under-specified queries, such as those that are ambiguous or multi-faceted. Existing approaches often rely on massive query logs and interaction data to generate a variety of possible query intents, which then can be used to re-rank documents. However, relying on user interaction data is problematic because one first needs a massive user base to build a sufficient log; public query logs are insufficient on their own. Given the recent success of causal language models (such as the Text-To-Text Transformer (T5) model) at text generation tasks, we explore the capacity of these models to generate potential query intents. We find that to encourage diversity in the generated queries, it is beneficial to adapt the model by including a new Distributional Causal Language Modeling (DCLM) objective during fine-tuning and a representation replacement during inference. Across six standard evaluation benchmarks, we find that our method (which we call IntenT5) improves search result diversity and attains (and sometimes exceeds) the diversity obtained when using query suggestions based on a proprietary query log. Our analysis shows that our approach is most effective for multi-faceted queries and is able to generalize effectively to queries that were unseen in training data.
Technology-assisted review (TAR) refers to iterative active learning workflows for document review in high recall retrieval (HRR) tasks. TAR research and most commercial TAR software have applied linear models such as logistic regression or support vector machines to lexical features. Transformer-based models with supervised tuning have been found to improve effectiveness on many text classification tasks, suggesting their use in TAR. We indeed find that the pre-trained BERT model reduces review volume by 30% in TAR workflows simulated on the RCV1-v2 newswire collection. In contrast, we find that linear models outperform BERT for simulated legal discovery topics on the Jeb Bush e-mail collection. This suggests the match between transformer pre-training corpora and the task domain is more important than generally appreciated. Additionally, we show that just-right language model fine-tuning on the task collection before starting active learning is critical. Both too little or too much fine-tuning results in performance worse than that of linear models, even for RCV1-v2.
Managing the data for Information Retrieval (IR) experiments can be challenging. Dataset documentation is scattered across the Internet and once one obtains a copy of the data, there are numerous different data formats to work with. Even basic formats can have subtle dataset-specific nuances that need to be considered for proper use. To help mitigate these challenges, we introduce a new robust and lightweight tool (ir_datases) for acquiring, managing, and performing typical operations over datasets used in IR. We primarily focus on textual datasets used for ad-hoc search. This tool provides both a python and command line interface to numerous IR datasets and benchmarks. To our knowledge, this is the most extensive tool of its kind. Integrations with popular IR indexing and experimentation toolkits demonstrate the tool's utility. We also provide documentation of these datasets through the ir_datasets catalog: https://ir-datasets.com/. The catalog acts as a hub for information on datasets used in IR, providing core information about what data each benchmark provides as well as links to more detailed information. We welcome community contributions and intend to continue to maintain and grow this tool.
Despite the recent successes of transformer-based models in terms of effectiveness on a variety of tasks, their decisions often remain opaque to humans. Explanations are particularly important for tasks like offensive language or toxicity detection on social media because a manual appeal process is often in place to dispute automatically flagged content. In this work, we propose a technique to improve the interpretability of these models, based on a simple and powerful assumption: a post is at least as toxic as its most toxic span. We incorporate this assumption into transformer models by scoring a post based on the maximum toxicity of its spans and augmenting the training process to identify correct spans. We find this approach effective and can produce explanations that exceed the quality of those provided by Logistic Regression analysis (often regarded as a highly-interpretable model), according to a human study.
Numerous studies have demonstrated the effectiveness of pretrained contextualized language models such as BERT and T5 for ad-hoc search. However, it is not well-understood why these methods are so effective, what makes some variants more effective than others, and what pitfalls they may have. We present a new comprehensive framework for Analyzing the Behavior of Neural IR ModeLs (ABNIRML), which includes new types of diagnostic tests that allow us to probe several characteristics---such as sensitivity to word order---that are not addressed by previous techniques. To demonstrate the value of the framework, we conduct an extensive empirical study that yields insights into the factors that contribute to the neural model's gains, and identify potential unintended biases the models exhibit. We find evidence that recent neural ranking models have fundamentally different characteristics from prior ranking models. For instance, these models can be highly influenced by altered document word order, sentence order and inflectional endings. They can also exhibit unexpected behaviors when additional content is added to documents, or when documents are expressed with different levels of fluency or formality. We find that these differences can depend on the architecture and not just the underlying language model.
With worldwide concerns surrounding the Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2), there is a rapidly growing body of scientific literature on the virus. Clinicians, researchers, and policy-makers need to be able to search these articles effectively. In this work, we present a zero-shot ranking algorithm that adapts to COVID-related scientific literature. Our approach filters training data from another collection down to medical-related queries, uses a neural re-ranking model pre-trained on scientific text (SciBERT), and filters the target document collection. This approach ranks top among zero-shot methods on the TREC COVID Round 1 leaderboard, and exhibits a P@5 of 0.80 and an nDCG@10 of 0.68 when evaluated on both Round 1 and 2 judgments. Despite not relying on TREC-COVID data, our method outperforms models that do. As one of the first search methods to thoroughly evaluate COVID-19 search, we hope that this serves as a strong baseline and helps in the global crisis.