Gabriella Kazai

In search of quality in crowdsourcing for search engine evaluation

Crowdsourcing is increasingly looked upon as a feasible alternative to traditional methods of gathering relevance labels for the evaluation of search engines, offering a solution to the scalability problem that hinders traditional approaches. However, crowdsourcing raises a range of questions regarding the quality of the resulting data. What indeed can be said about the quality of the data that is contributed by anonymous workers who are only paid cents for their efforts? Can higher pay guarantee better quality? Do better qualified workers produce higher quality labels? In this paper, we investigate these and similar questions via a series of controlled crowdsourcing experiments where we vary pay, required effort and worker qualifications and observe their effects on the resulting label quality, measured based on agreement with a gold set.

Crowdsourcing for book search evaluation: impact of hit design on comparative system ranking

The evaluation of information retrieval (IR) systems over special collections, such as large book repositories, is out of reach of traditional methods that rely upon editorial relevance judgments. Increasingly, the use of crowdsourcing to collect relevance labels has been regarded as a viable alternative that scales with modest costs. However, crowdsourcing suffers from undesirable worker practices and low quality contributions. In this paper we investigate the design and implementation of effective crowdsourcing tasks in the context of book search evaluation. We observe the impact of aspects of the Human Intelligence Task (HIT) design on the quality of relevance labels provided by the crowd. We assess the output in terms of label agreement with a gold standard data set and observe the effect of the crowdsourced relevance judgments on the resulting system rankings. This enables us to observe the effect of crowdsourcing on the entire IR evaluation process. Using the test set and experimental runs from the INEX 2010 Book Track, we find that varying the HIT design, and the pooling and document ordering strategies leads to considerable differences in agreement with the gold set labels. We then observe the impact of the crowdsourced relevance label sets on the relative system rankings using four IR performance metrics. System rankings based on MAP and Bpref remain less affected by different label sets while the Precision@10 and nDCG@10 lead to dramatically different system rankings, especially for labels acquired from HITs with weaker quality controls. Overall, we find that crowdsourcing can be an effective tool for the evaluation of IR systems, provided that care is taken when designing the HITs.

Community-based bayesian aggregation models for crowdsourcing

This paper addresses the problem of extracting accurate labels from crowdsourced datasets, a key challenge in crowdsourcing. Prior work has focused on modeling the reliability of individual workers, for instance, by way of confusion matrices, and using these latent traits to estimate the true labels more accurately. However, this strategy becomes ineffective when there are too few labels per worker to reliably estimate their quality. To mitigate this issue, we propose a novel community-based Bayesian label aggregation model, CommunityBCC, which assumes that crowd workers conform to a few different types, where each type represents a group of workers with similar confusion matrices. We assume that each worker belongs to a certain community, where the workers confusion matrix is similar to (a perturbation of) the communitys confusion matrix. Our model can then learn a set of key latent features: (i) the confusion matrix of each community, (ii) the community membership of each user, and (iii) the aggregated label of each item. We compare the performance of our model against established aggregation methods on a number of large-scale, real-world crowdsourcing datasets. Our experimental results show that our CommunityBCC model consistently outperforms state-of-the-art label aggregation methods, requiring, on average, 50% less data to pass the 90% accuracy mark.

An analysis of human factors and label accuracy in crowdsourcing relevance judgments

Crowdsourcing relevance judgments for the evaluation of search engines is used increasingly to overcome the issue of scalability that hinders traditional approaches relying on a fixed group of trusted expert judges. However, the benefits of crowdsourcing come with risks due to the engagement of a self-forming group of individuals—the crowd, motivated by different incentives, who complete the tasks with varying levels of attention and success. This increases the need for a careful design of crowdsourcing tasks that attracts the right crowd for the given task and promotes quality work. In this paper, we describe a series of experiments using Amazon’s Mechanical Turk, conducted to explore the ‘human’ characteristics of the crowds involved in a relevance assessment task. In the experiments, we vary the level of pay offered, the effort required to complete a task and the qualifications required of the workers. We observe the effects of these variables on the quality of the resulting relevance labels, measured based on agreement with a gold set, and correlate them with self-reported measures of various human factors. We elicit information from the workers about their motivations, interest and familiarity with the topic, perceived task difficulty, and satisfaction with the offered pay. We investigate how these factors combine with aspects of the task design and how they affect the accuracy of the resulting relevance labels. Based on the analysis of 960 HITs and 2,880 HIT assignments resulting in 19,200 relevance labels, we arrive at insights into the complex interaction of the observed factors and provide practical guidelines to crowdsourcing practitioners. In addition, we highlight challenges in the data analysis that stem from the peculiarity of the crowdsourcing environment where the sample of individuals engaged in specific work conditions are inherently influenced by the conditions themselves.

INEX 2007 Evaluation Measures

This paper describes the official measures of retrieval effectiveness that are employed for the Ad Hoc Track at INEX 2007. Whereas in earlier years all, but only, XML elements could be retrieved, the result format has been liberalized to arbitrary passages. In response, the INEX 2007 measures are based on the amount of highlighted text retrieved, leading to natural extensions of the well-established measures of precision and recall. The following measures are defined: The Focused Task is evaluated by interpolated precision at 1% recall (iP[0.01]) in terms of the highlighted text retrieved. The Relevant in Context Task is evaluated by mean average generalized precision (MAgP) where the generalized score per article is based on the retrieved highlighted text. The Best in Context Task is also evaluated by mean average generalized precision (MAgP) but here the generalized score per article is based on the distance to the assessors best-entry point.

On aggregating labels from multiple crowd workers to infer relevance of documents

We consider the problem of acquiring relevance judgements for information retrieval (IR) test collections through crowdsourcing when no true relevance labels are available. We collect multiple, possibly noisy relevance labels per document from workers of unknown labelling accuracy. We use these labels to infer the document relevance based on two methods. The first method is the commonly used majority voting (MV) which determines the document relevance based on the label that received the most votes, treating all the workers equally. The second is a probabilistic model that concurrently estimates the document relevance and the workers accuracy using expectation maximization (EM). We run simulations and conduct experiments with crowdsourced relevance labels from the INEX 2010 Book Search track to investigate the accuracy and robustness of the relevance assessments to the noisy labels. We observe the effect of the derived relevance judgments on the ranking of the search systems. Our experimental results show that the EM method outperforms the MV method in the accuracy of relevance assessments and IR systems ranking. The performance improvements are especially noticeable when the number of labels per document is small and the labels are of varied quality.

Social book search: comparing topical relevance judgements and book suggestions for evaluation

The Web and social media give us access to a wealth of information, not only different in quantity but also in character---traditional descriptions from professionals are now supplemented with user generated content. This challenges modern search systems based on the classical model of topical relevance and ad hoc search: How does their effectiveness transfer to the changing nature of information and to the changing types of information needs and search tasks? We use the INEX 2011 Books and Social Search Tracks collection of book descriptions from Amazon and social cataloguing site LibraryThing. We compare classical IR with social book search in the context of the LibraryThing discussion forums where members ask for book suggestions. Specifically, we compare book suggestions on the forum with Mechanical Turk judgements on topical relevance and recommendation, both the judgements directly and their resulting evaluation of retrieval systems. First, the book suggestions on the forum are a complete enough set of relevance judgements for system evaluation. Second, topical relevance judgements result in a different system ranking from evaluation based on the forum suggestions. Although it is an important aspect for social book search, topical relevance is not sufficient for evaluation. Third, professional metadata alone is often not enough to determine the topical relevance of a book. User reviews provide a better signal for topical relevance. Fourth, user-generated content is more effective for social book search than professional metadata. Based on our findings, we propose an experimental evaluation that better reflects the complexities of social book search.

Wikipedia pages as entry points for book search

A lot of the worlds knowledge is stored in books, which, as a result of recent mass-digitisation efforts, are increasingly available online. Search engines, such as Google Books, provide mechanisms for searchers to enter this vast knowledge space using queries as entry points. In this paper, we view Wikipedia as a summary of this world knowledge and aim to use this resource to guide users to relevant books. Thus, we investigate possible ways of using Wikipedia as an intermediary between the users query and a collection of books being searched. We experiment with traditional query expansion techniques, exploiting Wikipedia articles as rich sources of information that can augment the users query. We then propose a novel approach based on link distance in an extended Wikipedia graph: we associate books with Wikipedia pages that cite these books and use the link distance between these nodes and the pages that match the user query as an estimation of a books relevance to the query. Our results show that a) classical query expansion using terms extracted from query pages leads to increased precision, and b) link distance between query and book pages in Wikipedia provides a good indicator of relevance that can boost the retrieval score of relevant books in the result ranking of a book search engine.

An analysis of systematic judging errors in information retrieval

Test collections are powerful mechanisms for the evaluation and optimization of information retrieval systems. However, there is reported evidence that experiment outcomes can be affected by changes to the judging guidelines or changes in the judge population. This paper examines such effects in a web search setting, comparing the judgments of four groups of judges: NIST Web Track judges, untrained crowd workers and two groups of trained judges of a commercial search engine. Our goal is to identify systematic judging errors by comparing the labels contributed by the different groups, working under the same or different judging guidelines. In particular, we focus on detecting systematic differences in judging depending on specific characteristics of the queries and URLs. For example, we ask whether a given population of judges, working under a given set of judging guidelines, are more likely to consistently overrate Wikipedia pages than another group judging under the same instructions. Our approach is to identify judging errors with respect to a consensus set, a judged gold set and a set of user clicks. We further demonstrate how such biases can affect the training of retrieval systems.

Setting up a competition framework for the evaluation of structure extraction from OCR-ed books

This paper describes the setup of the Book Structure Extraction competition run at ICDAR 2009. The goal of the competition was to evaluate and compare automatic techniques for deriving structure information from digitized books, which could then be used to aid navigation inside the books. More specifically, the task that participants faced was to construct hyperlinked tables of contents for a collection of 1,000 digitized books. This paper describes the setup of the competition and its challenges. It introduces and discusses the book collection used in the task, the collaborative construction of the ground truth, the evaluation measures, and the evaluation results. The paper also introduces a data set to be used freely for research evaluation purposes.