Search

Within many domains, such as news, medicine and patent, documents contain a variety of fields such as title, author, body, source, etc. As such fielded retrieval models that query across fields are often employed. It is largely presumed that fielding provides a better representation of the document and offers more control when querying, and that this will lead to improved retrieval performance. However, depending on how the fields are weighted and if the fields are populated, the retrieval algorithm may unduly favour certain documents over others. This is known as algorithmic bias and it may be detrimental to retrieval systems performance. In this paper, we explore the impact of fielding on retrieval bias and performance across a variety of TREC News Test Collections. We perform an extensive large-scale analysis on two types of fielded retrieval model variations that are based on the popular BM25 retrieval algorithm where either: fields are scored independently and then combined (Model 1), or fields are first combined and then scored (Model 2). Our findings show that for Model 1 fielding, a strong correlation exists between retrieval bias and performance such that as title fields are weighted more heavily, bias increases, while retrieval performance decreases. When weighting is applied to content-based fields, performance increases as bias decreases, showing that relying more on content may be favourable in terms of fairness and performance. On the other hand, for Model 2 fielding, the relationship between retrieval bias and performance is more complex. But, crucially we show that Model 2 fielding results in lower retrieval bias and greater performance than Model 1 fielding. And, we observed that under Model 1, news articles without titles are substantially less retrievable (i.e. more susceptible to algorithmic bias). These findings have serious ramifications as many popular Open Source Information Retrieval frameworks, commonly used by professional searchers, use the default implementation of Model 1 for their fielded search capability. This research shows the importance of analysing retrieval algorithms with respect to both bias and performance to ensure they minimize any unwanted or unintended biases when maximising performance. Further work is required to examine this phenomenon in more detail and to design fielded retrieval models that have the advantages of control and performance without detrimental biases.

Algorithmic bias presents a difficult challenge within Information Retrieval. Long has it been known that certain algorithms favour particular documents due to attributes of these documents that are not directly related to relevance. The evaluation of bias has recently been made possible through the use of retrievability, a quantifiable measure of bias. While evaluating bias is relatively novel, the evaluation of performance has been common since the dawn of the Cranfield approach and TREC. To evaluate performance, a pool of documents to be judged by human assessors is created from the collection. This pooling approach has faced accusations of bias due to the fact that the state of the art algorithms were used to create it, thus the inclusion of biases associated with these algorithms may be included in the pool. The introduction of retrievability has provided a mechanism to evaluate the bias of these pools. This work evaluates the varying degrees of bias present in the groups of relevant and non-relevant documents for topics. The differentiating power of a system is also evaluated by examining the documents from the pool that are retrieved for each topic. The analysis finds that the systems that perform better, tend to have a higher chance of retrieving a relevant document rather than a non-relevant document for a topic prior to retrieval, indicating that retrieval systems which perform better at TREC are already predisposed to agree with the judgements regardless of the query posed.

Query expansion is a useful retrieval mechanism for creating more verbose queries from the users initial keyword search. Query expansion generally have multiple parameters that allowthe user to define how many terms and where those terms come from are introduced to the expanded query. However, the idea that query expansion may be introducing biases into the system by selecting terms from overly retrievable documents has never been formally evaluated. In this work, the relationship between performance and retrievability bias is explored when various query expansion methods are employed to aide retrieval. Several parameters are altered, independently, to identify those that have an impact on bias. Parameters altered include; Rocchio's beta, length normalisation parameters, the number of terms added and the number of documents those terms are extracted from. The evaluation performed here identifies a strong correlation between performance and retrievability bias, suggesting that performance is increased by making the system more biased thus more likely to pick terms from a set of overly retrievable documents.

Retrievability is an independent evaluation measure that offers insights to an aspect of retrieval systems that performance and efficiency measures do not. Retrievability is often used to calculate the retrievability bias, an indication of how accessible a system makes all the documents in a collection. Generally, computing the retrievability bias of a system requires a colossal number of queries to be issued for the system to gain an accurate estimate of the bias. However, it is often the case that the accuracy of the estimate is not of importance, but the relationship between the estimate of bias and performance when tuning a systems parameters. As such, reaching a stable estimation of bias for the system is more important than getting very accurate retrievability scores for individual documents. This work explores the idea of using topical subsets of the collection for query generation and bias estimation to form a local estimate of bias which correlates with the global estimate of retrievability bias. By using topical subsets, it would be possible to reduce the volume of queries required to reach an accurate estimate of retrievability bias, reducing the time and resources required to perform a retrievability analysis. Findings suggest that this is a viable approach to estimating retrievability bias and that the number of queries required can be reduced to less than a quarter of what was previously thought necessary.