Evaluating Relevance Judgments with Pairwise Discriminative Power

Judging the relevance of documents for an information need is an activity that underpins the most widely-used approach in the evaluation of information retrieval systems. In this study we investigate the relationship between how long it takes an assessor to judge document relevance, and three key factors that may influence the judging scenario: the difficulty of the search topic for which relevance is being assessed; the degree to which the documents are relevant to the search topic; and, the order in which the documents are presented for judging. Two potential confounding influences on judgment speed are differences in individual reading ability, and the length of documents that are being assessed. We therefore propose two measures to investigate the above factors: normalized processing speed (NPS), which adjusts the number of words that were processed per minute by taking into account differences in reading speed between judges, and normalized dwell time (NDT), which adjusts the duration that a judge spent reading a document relative to document length. Note that these two measures have different relationships with overall judgment speed: a direct relationship for NPS, and an inverse relationship for NDT. The results of a small-scale user study show a statistically significant relationship between judgment speed and topic difficulty: for easier topics, assessors process more quickly (higher NPS), and spend less time overall (lower NDT). There is also a statistically significant relationship between the level of relevance of the document being assessed and overall judgment speed, with assessors taking less time for non-relevant documents. Finally, our results suggest that the presentation order of documents can also affect overall judgment speed, with assessors spending less time (smaller NDT) when documents are presented in relevance order than docID order. However, these ordering effects are not significant when also accounting for document length variance (NPS).

Test collection is used to evaluate the information retrieval systems in laboratory-based evaluation experimentation. In a classic setting, generating relevance judgments involves human assessors and is a costly and time consuming task. Researchers and practitioners are still being challenged in performing reliable and low-cost evaluation of retrieval systems. Crowdsourcing as a novel method of data acquisition is broadly used in many research fields. It has been proven that crowdsourcing is an inexpensive and quick solution as well as a reliable alternative for creating relevance judgments. One of the crowdsourcing applications in IR is to judge relevancy of query document pair. In order to have a successful crowdsourcing experiment, the relevance judgment tasks should be designed precisely to emphasize quality control. This paper is intended to explore different factors that have an influence on the accuracy of relevance judgments accomplished by workers and how to intensify the reliability of judgments in crowdsourcing experiment.

Document corpora are key components of information retrieval test collections. However, for certain tasks, such as evaluating the effectiveness of a new retrieval technique or estimating the parameters of a learning to rank model, a corpus alone is not enough. For these tasks, queries and relevance judgments associated with the corpus are also necessary. However, researchers often find themselves in scenarios where they only have access to a corpus, in which case evaluation and learning to rank become challenging. Document corpora are relatively straightforward to gather. On the other hand, obtaining queries and relevance judgments for a given corpus is costly. In production environments, it may be possible to obtain low-cost relevance information using query and click logs. However, in more constrained research environments these options are not available, and relevance judgments are usually provided by humans. To reduce the cost of this potentially expensive process, researchers have developed low-cost evaluation strategies, including minimal test collections [2] and crowdsourcing [1]. Despite the usefulness of these strategies, the resulting relevance judgments cannot easily be “ported” to a new or different corpus. To overcome these issues, we propose a new method to reduce manual annotation costs by transferring relevance judgments across corpora. Assuming that a set of queries and relevance judgments have been manually constructed for a source document corpus Ds, our goal is to automatically construct a test collection for a target document corpus Dt by projecting the existing test collection from Ds onto Dt. The goal of projecting test collections is not to produce manual quality test collections. In fact, it is assumed that projected test collections will contain noisy relevance judgments (i.e., ones which humans are unlikely to agree with). The important question, however, is whether these noisy projected judgments are useful for training ranking models in the target corpus.

Recent studies in information retrieval have shown that gender biases have found their way into representational and algorithmic aspects of computational models. In this paper, we focus specifically on gender biases in information retrieval gold standard datasets, often referred to as relevance judgements. While not explored in the past, we submit that it is important to understand and measure the extent to which gender biases may be presented in information retrieval relevance judgements primarily because relevance judgements are not only the primary source for evaluating IR techniques but are also widely used for training end-to-end neural ranking methods. As such, the presence of bias in relevance judgements would immediately find its way into how retrieval methods operate in practice. Based on a fine-tuned BERT model, we show how queries can be labelled for gender at scale based on which we label MS MARCO queries. We then show how different psychological characteristics are exhibited within documents associated with gendered queries within the relevance judgement datasets. Our observations show that stereotypical biases are prevalent in relevance judgement documents.

From highly relevant to not relevant: examining different regions of relevance

Despite the popularity of Twitter for research, there are very few publicly available corpora, and those which are available are either too small or unsuitable for tasks such as event detection. This is partially due to a number of issues associated with the creation of Twitter corpora, including restrictions on the distribution of the tweets and the difficultly of creating relevance judgements at such a large scale. The difficulty of creating relevance judgements for the task of event detection is further hampered by ambiguity in the definition of event. In this paper, we propose a methodology for the creation of an event detection corpus. Specifically, we first create a new corpus that covers a period of 4 weeks and contains over 120 million tweets, which we make available for research. We then propose a definition of event which fits the characteristics of Twitter, and using this definition, we generate a set of relevance judgements aimed specifically at the task of event detection. To do so, we make use of existing state-of-the-art event detection approaches and Wikipedia to generate a set of candidate events with associated tweets. We then use crowdsourcing to gather relevance judgements, and discuss the quality of results, including how we ensured integrity and prevented spam. As a result of this process, along with our Twitter corpus, we release relevance judgements containing over 150,000 tweets, covering more than 500 events, which can be used for the evaluation of event detection approaches.

Evaluation is crucial in assessing the effectiveness of new information retrieval and human computer interaction techniques and systems. Relevance judgements are often performed by humans, which makes obtaining them expensive and time consuming. Consequently, relevance judgements are usually performed only on a subset of a given collection of data or experimental results with a focus on the top ranked documents. However, when assessing the performance of exploratory search systems, the diversity or subjective relevance of documents that the user was presented with over a search session are often of more importance than the relative ranking of top documents. In order to perform these types of assessment, all the documents in a given collection need to be judged for relevance. In this paper, we propose an approach based on topic modeling that can greatly accelerate document relevance judgment of an entire document collection with an expert assessor needing to mark only a small subset of documents from a given collection. Experimental results show a substantial overlap between relevance judgments compared to a human assessor.

Eye movement data can provide an in‐depth view of human reasoning and the decision‐making process, and modern information retrieval (IR) research can benefit from the analysis of this type of data. The aim of this research was to examine the relationship between relevance criteria use and visual behavior in the context of predictive relevance judgments. To address this objective, a multimethod research design was employed that involved observation of participants’ eye movements, talk‐aloud protocols, and postsearch interviews. Specifically, the results reported in this article came from the analysis of 281 predictive relevance judgments made by 24 participants using the Google search engine. We present a novel stepwise methodological framework for the analysis of relevance judgments and eye movements on the Web and show new patterns of relevance criteria use during predictive relevance judgment. For example, the findings showed an effect of ranking order and surrogate components (Title, Summary, and URL) on the use of relevance criteria. Also, differences were observed in the cognitive effort spent between very relevant and not relevant judgments. We conclude with the implications of this study for IR research.

Effect of verbal comprehension skill and self-reported features on reliability of crowdsourced relevance judgments

Reliable evaluation of Information Retrieval systems requires large amounts of relevance judgments. Making these annotations is quite complex and tedious for many Music Information Retrieval tasks, so performing such evaluations requires too much effort. A low-cost alternative is the application of Minimal Test Collection algorithms, which offer quite reliable results while significantly reducing the annotation effort. The idea is to incrementally select what documents to judge so that we can compute estimates of the effectiveness differences between systems with a certain degree of confidence. In this paper we show a first approach towards its application to the evaluation of the Audio Music Similarity and Retrieval task, run by the annual MIREX evaluation campaign. An analysis with the MIREX 2011 data shows that the judging effort can be reduced to about 35% to obtain results with 95% confidence.

Relevance is an essential concept in Information Retrieval (IR). Recent studies using brain imaging have significantly contributed towards the understanding of this concept, but only as a binary notion, i.e. a document being judged as relevant or non-relevant. While such a binary division is prevalent in IR, seminal theories have proposed relevance as a graded variable; i.e. having different degrees. In this paper, we aim to investigate the brain activity associated with relevance when it is treated as a graded concept. Twenty-five participants provided graded relevance judgements in the context of a Question Answering (Q/A) Task, during assessment with an electroencephalogram (EEG). Our findings show that significant differences in event-related potentials (ERPs) were observed in response to information segments processed in the context of high-relevance, low-relevance and no-relevance, supporting the concept of graded relevance. We speculate that differences in attentional engagement, semantic mismatch (between the question and answer) and memory processing underpin the electrophysiological responses to the graded relevance judgements. We believe our conclusions constitute an important step in unravelling the nature of graded relevance and knowledge of the electrophysiological modulation to each grade of relevance will help to improve the design and evaluation of IR systems.

Analyzing and modeling user behavior in the Information Retrieval (IR) process is important for IR research. In this paper, we propose a new research schema in IR research. In this schema, we first investigate humans' cognitive process when completing a specific IR task and build cognitive models for that task. The findings in the investigation and the proposed cognitive models are then utilized to improve machines' performance in the IR task. Through this research schema, we revisit three IR tasks. In the first study, we carefully analyze users' clicking behavior in mobile search and propose a mobile click model to extract relevance feedback from the mobile search logs. In the second study, we conduct an eye-tracking study to investigate how human assessors read a document during relevance judgment task and adopt the findings in building a novel retrieval model that can better approximate humans' relevance judgment. In the last study, we conduct another eye-tracking study to investigate humans' reading behavior when completing the reading comprehension task. We build a prediction model for user attention and leverage the predicted attention signals to improve the machine reading comprehension model. By successfully adopting this research schema to three different IR tasks, we demonstrate its effectiveness and generalizability.

Research in Information Retrieval has progressed against a background of rapidly increasing corpus size and heterogeneity, with every advance in technology quickly followed by a desire to organize and search more unstructured, more heterogeneous, and even bigger corpora. But as retrieval problems get larger and more complicated, evaluating the ranking performance of a retrieval engine gets harder: evaluation requires human judgments of the relevance of documents to queries, and for very large corpora the cost of acquiring these judgments may be insurmountable. This cost limits the types of problems researchers can study as well as the data they can be studied on. We present methods for understanding performance differences between retrieval engines in the presence of missing and noisy relevance judgments. The work introduces a model of the cost of experimentation that incorporates the cost of human judgments as well as the cost of drawing incorrect conclusions about differences between engines in both the training and testing phases of engine development. Through adopting a view of evaluation that is more concerned with distributions over performance differences rather than estimates of absolute performance, the expected cost can be minimized so as to reliably differentiate between engines with less than 1% of the human effort that has been used in past experiments.

Network Intrusion Detection Systems (NIDSs) are an increasingly important tool for the prevention and mitigation of cyber attacks. Over the past years, a lot of research efforts have aimed at leveraging the increasingly powerful models of Machine Learning (ML) for this purpose. A number of labelled synthetic datasets have been generated and made publicly available by researchers, and they have become the benchmarks via which new ML-based NIDS classifiers are being evaluated. Recently published results show excellent classification performance with these datasets, increasingly approaching 100 percent performance across key evaluation metrics such as Accuracy, F1 score, AUC, etc. Unfortunately, we have not yet seen these excellent academic research results translated into practical NIDS systems with such near-perfect performance. This motivated our research presented in this paper, where we analyse the statistical properties of the benign traffic in three of the more recent and relevant NIDS datasets, (CIC_IDS, UNSW_NB15, TON_IOT), by converting them into a common flow format. As a comparison, we consider two datasets obtained from real-world production networks, one from a university network and one from a medium size Internet Service Provider (ISP). Our results show that the two real-world datasets are quite similar among themselves in regards to most of the considered statistical features. Equally, the three synthetic datasets are also relatively similar within their group. However, and most importantly, our results show a distinct difference of most of the considered statistical features between the three synthetic datasets and the two real-world datasets. Since ML relies on the basic assumption of training and test datasets being sampled from the same distribution, this raises the question of how well the performance results of ML-classifiers trained on the considered synthetic datasets can translate and generalise to real-world networks. We believe this is an interesting and relevant question which provides motivation for further research in this space.

Multichannel communication has great potential in environments with unknown interference patterns. However, existing link quality metrics and models are generally established and verified under a single-channel scenario, which does not consider the impacts of radio interference and channel change. Therefore, it is hard to directly judge whether these metrics and models are still valid under a multichannel scenario. This paper empirically analyzes the applicability of popular link quality metrics and models in multiple channels with different interference levels. Results show that the link quality estimation (LQE) capability of traditional metrics will be affected by the interference level of the channel, which makes the conclusions obtained under a single-channel scenario no longer valid. Meanwhile, traditional LQE models are basically not adaptive to radio interference and channel change. They are only valid for channels with similar interference under which they are modeled. If these models are directly used under a multichannel scenario, the link quality will be overestimated inevitably. In other words, traditional LQE metrics and models cannot be directly used in the multichannel scenario. It is necessary to deeply analyze the statistical characteristics of popular link quality metrics in multiple typical channels and design channel and interference adaptive metrics and models to support effective multichannel communication.