Search

Council members and policy workers need to understand (long-term) processes that lead to decisions. Gaining such an overview of a topic through a search engine can be challenging however, as searching a complex topic can result in an overwhelming number of documents and does not show how these documents are interrelated. This study investigates how to create an overview of a decision-making process, which may be integrated into a search engine. Interviews show that policy workers consider documents relevant to the overview when the document and proposal were both created in response to the same council decision document. We identify such provenance based on the co-citation of documents and textual references between documents. In an exploratory user study, policy workers are tasked to understand the development of policy proposals based on provided timelines. Their relevance assessments show that our approach nearly exclusively finds relevant documents (a precision of 0.97). Whereas the proposed approach identifies 91% of references made in documents, it only finds an exact target document in 39% of the total references. A further 52% of references finds a subset of documents including the target. A human in the loop can aid in finding the exact documents, and potentially add documents based on their domain expertise. The proposed approach creates an overview of a city council’s decision-making process on a given topic with high precision, and might apply to other domains oriented around a decision-making process.

Expert search systems help professionals find colleagues with specific expertise. Expert search results can be presented as a list of documents with their associated experts, or as a list of candidate experts with evidence for their expertise based on documents they authored. The type of result may affect search behaviour, and therefore search task performance. Previous work has not considered such effects from the result presentation, focusing instead on how to rank experts or on ways to interact with the search results. We compare the task performance of novice users using either a document-centric interface (where each search result is a document and its associated expert) or a candidate-centric interface (where each search result is a candidate expert and their associated documents). We also compare candidate-centric and document-centric ranking functions per interface. A post-experiment survey indicated that two variables affect which interface participants preferred: the retrieval unit (candidates or documents) and the complexity (number of documents per search result). These variables affected participants' search strategy, and consequently their task performance. A quantitative analysis revealed that 1) using the candidate-centric interface results in a higher rate of correctly completed tasks, as users evaluate candidates more thoroughly, and 2) the document-centric ranking yields faster task completion. Weak evidence of a statistical interaction effect was found that prevents a straightforward combination of the most effective interface type and the most efficient ranking type. Present work resulted in a more effective, albeit less efficient, search engine for expert search at the municipality of Utrecht.

We adapt previous literature on search tasks for developing a domain-specific search engine that supports the search tasks of policy workers. To characterise the search tasks we conducted two rounds of interviews with policy workers at the municipality of Utrecht, and found that they face different challenges depending on the complexity of the task. During simple tasks, policy workers face information overload and time pressures, especially during web-based searches. For complex tasks, users prefer finding domain experts within their organisation to obtain the necessary information, which requires a different type of search functionality. To support simple tasks, we developed a web search engine that indexes web pages from authoritative sources only. We tested the hypothesis that users prefer expert search over web search for complex tasks and found that supporting complex tasks requires integrating functionality that enables finding internal experts within the broader web search engine. We constructed representative tasks to evaluate the proposed system’s effectiveness and efficiency, and found that it improved user performance. The search functionality developed could be standardised for use by policy workers in different municipalities within the Netherlands.

Professional search is a problem area in which many facets of information retrieval are addressed, both system-related (e.g. distributed search) and user-related (e.g. complex information needs), and the interface between user and system (e.g. supporting exploratory search tasks). Professional search tasks have specific requirements, different from the requirements of generic web search. The aim of this workshop is to bring together researchers to work on the requirements and challenges of professional search from different angles. We will have an interactive workshop where researchers not only present their scientific results but also work together on the definition of future challenges and solutions with input from information professionals. The workshop will deliver a roadmap of research directions for the years to come.

A Web archive usually contains multiple versions of documents crawled from the Web at different points in time. One possible way for users to access a Web archive is through full-text search systems. However, previous studies have shown that these systems can induce a bias, known as the retrievability bias, on the accessibility of documents in community-collected collections (such as TREC collections). This bias can be measured by analyzing the distribution of the retrievability scores for each document in a collection, quantifying the likelihood of a document’s retrieval. We investigate the suitability of retrievability scores in retrieval systems that consider every version of a document in a Web archive as an independent document. We show that the retrievability of documents can vary for different versions of the same document and that retrieval systems induce biases to different extents. We quantify this bias for a retrieval system which is adapted to handle multiple versions of the same document. The retrieval system indexes each version of a document independently, and we refine the search results using two techniques to aggregate similar versions. The first approach is to collapse similar versions of a document based on content similarity. The second approach is to collapse all versions of the same document based on their URLs. In both cases, we found that the degree of bias is related to the aggregation level of versions of the same document. Finally, we study the effect of bias across time using the retrievability measure. Specifically, we investigate whether the number of documents crawled in a particular year correlates with the number of documents in the search results from that year. Assuming queries are not inherently temporal in nature, the analysis is based on the timestamps of documents in the search results returned using the retrieval model for all queries. The results show a relation between the number of documents per year and the

Professional search is a problem area in which many facets of information retrieval are addressed, both system-related (e.g. distributed search) and user-related (e.g. complex information needs), and the interface between user and system (e.g. supporting exploratory search tasks). Professional search tasks have specific requirements, different from the requirements of generic web search. The aim of this workshop is to bring together researchers to work on the requirements and challenges of professional search from different angles. We will have an interactive workshop where researchers not only present their scientific results but also work together on the definition of future challenges and solutions with input from information professionals. The workshop will deliver a roadmap of research directions for the years to come.

A Web archive usually contains multiple versions of documents crawled from the Web at different points in time. One possible way for users to access a Web archive is through full-text search systems. However, previous studies have shown that these systems can induce a bias, known as the retrievability bias, on the accessibility of documents in community-collected collections (such as TREC collections). This bias can be measured by analyzing the distribution of the retrievability scores for each document in a collection, quantifying the likelihood of a document’s retrieval. We investigate the suitability of retrievability scores in retrieval systems that consider every version of a document in a Web archive as an independent document. We show that the retrievability of documents can vary for different versions of the same document and that retrieval systems induce biases to different extents. We quantify this bias for a retrieval system which is adapted to handle multiple versions of the same document. The retrieval system indexes each version of a document independently, and we refine the search results using two techniques to aggregate similar versions. The first approach is to collapse similar versions of a document based on content similarity. The second approach is to collapse all versions of the same document based on their URLs. In both cases, we found that the degree of bias is related to the aggregation level of versions of the same document. Finally, we study the effect of bias across time using the retrievability measure. Specifically, we investigate whether the number of documents crawled in a particular year correlates with the number of documents in the search results from that year. Assuming queries are not inherently temporal in nature, the analysis is based on the timestamps of documents in the search results returned using the retrieval model for all queries. The results show a relation between the number of documents per year and the

In the evaluation of recommender systems, the quality of recommendations made by a newly proposed algorithm is compared to the state-of-the-art, using a given quality measure and dataset. Validity of the evaluation depends on the assumption that the evaluation does not exhibit artefacts resulting from the process of collecting the dataset. The main difference between online and offline evaluation is that in the online setting, the user’s response to a recommendation is only observed once. We used the NewsREEL challenge to gain a deeper understanding of the implications of this difference for making comparisons between different recommender systems. The experiments aim to quantify the expected degree of variation in performance that cannot be attributed to differences between systems. We classify and discuss the non-algorithmic causes of performance differences observed.

Successful news recommendation requires facing the challenges of dynamic item sets, contextual item relevance, and of fulfilling non-functional requirements, such as response time. The CLEF News- REEL challenge is a campaign-style evaluation lab allowing participants to tackle news recommendation and to optimize and evaluate their recommender algorithms both online and offline. In this paper, we summarize the objectives and challenges of NewsREEL 2016. We cover two contrasting perspectives on the challenge: that of the operator (the business providing recommendations) and that of the challenge participant (the researchers developing recommender algorithms). In the intersection of these perspectives, new insights can be gained on how to effectively evaluate real-time stream recommendation algorithms.

We present a comparison of recommender systems algorithms along four dimensions. The first dimension is offline evaluation where we compare the performance of our algorithms in an offline setting. The second dimension is online evaluation where we deploy recommender algorithms online with a view to comparing their performance patterns. The third dimension is time, where we compare our algorithms in two different years: 2015 and 2016. The fourth dimension is the quantification of the effect of non-Algorithmic factors on the performance of an online recommender system by using an A/A test. We then analyze the performance similarities and differences along these dimensions in an attempt to draw meaningful patterns and conclusions.

Bias in the retrieval of documents can directly influence the information access of a digital library. In the worst case, systematic favoritism for a certain type of document can render other parts of the collection invisible to users. This potential bias can be evaluated by measuring the retrievability for all documents in a collection. Previous evaluations have been performed on TREC collections using simulated query sets. The question remains, however, how representative this approach is of more realistic settings. To address this question, we investigate the effectiveness of the retrievability measure using a large digitized newspaper corpus, featuring two characteristics that distinguishes our experiments from previous studies: (1) compared to TREC collections, our collection contains noise originating from OCR processing, historical spelling and use of language; and (2) instead of simulated queries, the collection comes with real user query logs including click data. First, we assess the retrievability bias imposed on the newspaper collection by different IR models. We assess the retrievability measure and confirm its ability to capture the retrievability bias in our setup. Second, we show how simulated queries differ from real user queries regarding term frequency and prevalence of named entities, and how this affects the retrievability results.

Web archives preserve the fast changing Web by repeatedly crawling its content. The crawling strategy has an influence on the data that is archived. We use link anchor text of two Web crawls created with different crawling strategies in order to compare their coverage of past popular topics. One of our crawls was collected by the National Library of the Netherlands (KB) using a depth-first strategy on manually selected websites from the .nl domain, with the goal to crawl websites as completes as possible. The second crawl was collected by the Common Crawl foundation using a breadth-first strategy on the entire Web, this strategy focuses on discovering as many links as possible. The two crawls differ in their scope of coverage, while the KB dataset covers mainly the Dutch domain, the Common Crawl dataset covers websites from the entire Web. Therefore, we used three different sources to identify topics that were popular on the Web; both at the global level (entire Web) and at the national level (.nl domain): Google Trends, WikiStats, and queries collected from users of the Dutch historic newspaper archive. The two crawls are different in terms of their size, number of included websites and domains. To allow fair comparison between the two crawls, we created sub-collections from the Common Crawl dataset based on the .nl domain and the KB seeds. Using simple exact string matching between anchor texts and popular topics from the three different sources, we found that the breadth-first crawl covered more topics than the depth-first crawl. Surprisingly, this is not limited to popular topics from the entire Web but also applies to topics that were popular in the .nl domain.

In the evaluation of recommender systems, the quality of recommendations made by a newly proposed algorithm is compared to the state-of-the-art, using a given quality measure and dataset. Validity of the evaluation depends on the assumption that the evaluation does not exhibit artefacts resulting from the process of collecting the dataset. The main difference between online and offline evaluation is that in the online setting, the user’s response to a recommendation is only observed once. We used the NewsREEL challenge to gain a deeper understanding of the implications of this difference for making comparisons between different recommender systems. The experiments aim to quantify the expected degree of variation in performance that cannot be attributed to differences between systems. We classify and discuss the non-algorithmic causes of performance differences observed.

Successful news recommendation requires facing the challenges of dynamic item sets, contextual item relevance, and of fulfilling non-functional requirements, such as response time. The CLEF News- REEL challenge is a campaign-style evaluation lab allowing participants to tackle news recommendation and to optimize and evaluate their recommender algorithms both online and offline. In this paper, we summarize the objectives and challenges of NewsREEL 2016. We cover two contrasting perspectives on the challenge: that of the operator (the business providing recommendations) and that of the challenge participant (the researchers developing recommender algorithms). In the intersection of these perspectives, new insights can be gained on how to effectively evaluate real-time stream recommendation algorithms.

We present a comparison of recommender systems algorithms along four dimensions. The first dimension is offline evaluation where we compare the performance of our algorithms in an offline setting. The second dimension is online evaluation where we deploy recommender algorithms online with a view to comparing their performance patterns. The third dimension is time, where we compare our algorithms in two different years: 2015 and 2016. The fourth dimension is the quantification of the effect of non-Algorithmic factors on the performance of an online recommender system by using an A/A test. We then analyze the performance similarities and differences along these dimensions in an attempt to draw meaningful patterns and conclusions.

Bias in the retrieval of documents can directly influence the information access of a digital library. In the worst case, systematic favoritism for a certain type of document can render other parts of the collection invisible to users. This potential bias can be evaluated by measuring the retrievability for all documents in a collection. Previous evaluations have been performed on TREC collections using simulated query sets. The question remains, however, how representative this approach is of more realistic settings. To address this question, we investigate the effectiveness of the retrievability measure using a large digitized newspaper corpus, featuring two characteristics that distinguishes our experiments from previous studies: (1) compared to TREC collections, our collection contains noise originating from OCR processing, historical spelling and use of language; and (2) instead of simulated queries, the collection comes with real user query logs including click data. First, we assess the retrievability bias imposed on the newspaper collection by different IR models. We assess the retrievability measure and confirm its ability to capture the retrievability bias in our setup. Second, we show how simulated queries differ from real user queries regarding term frequency and prevalence of named entities, and how this affects the retrievability results.

Web archives preserve the fast changing Web by repeatedly crawling its content. The crawling strategy has an influence on the data that is archived. We use link anchor text of two Web crawls created with different crawling strategies in order to compare their coverage of past popular topics. One of our crawls was collected by the National Library of the Netherlands (KB) using a depth-first strategy on manually selected websites from the .nl domain, with the goal to crawl websites as completes as possible. The second crawl was collected by the Common Crawl foundation using a breadth-first strategy on the entire Web, this strategy focuses on discovering as many links as possible. The two crawls differ in their scope of coverage, while the KB dataset covers mainly the Dutch domain, the Common Crawl dataset covers websites from the entire Web. Therefore, we used three different sources to identify topics that were popular on the Web; both at the global level (entire Web) and at the national level (.nl domain): Google Trends, WikiStats, and queries collected from users of the Dutch historic newspaper archive. The two crawls are different in terms of their size, number of included websites and domains. To allow fair comparison between the two crawls, we created sub-collections from the Common Crawl dataset based on the .nl domain and the KB seeds. Using simple exact string matching between anchor texts and popular topics from the three different sources, we found that the breadth-first crawl covered more topics than the depth-first crawl. Surprisingly, this is not limited to popular topics from the entire Web but also applies to topics that were popular in the .nl domain.

Web archives attempt to preserve the fast changing web, yet they will always be incomplete. Due to restrictions in crawling depth, crawling frequency, and restrictive selection policies, large parts of the Web are unarchived and, therefore, lost to posterity. In this paper, we propose an approach to uncover unarchived web pages and websites and to reconstruct different types of descriptions for these pages and sites, based on links and anchor text in the set of crawled pages. We experiment with this approach on the Dutch Web Archive and evaluate the usefulness of page and host-level representations of unarchived content. Our main findings are the following: First, the crawled web contains evidence of a remarkable number of unarchived pages and websites, potentially dramatically increasing the coverage of a Web archive. Second, the link and anchor text have a highly skewed distribution: popular pages such as home pages have more links pointing to them and more terms in the anchor text, but the richness tapers off quickly. Aggregating web page evidence to the host-level leads to significantly richer representations, but the distribution remains skewed. Third, the succinct representation is generally rich enough to uniquely identify pages on the unarchived web: in a known-item search setting we can retrieve unarchived web pages within the first ranks on average, with host-level representations leading to further improvement of the retrieval effectiveness for websites.

Cumulative Citation Recommendation (CCR) is defined as: given a stream of documents on one hand and Knowledge Base (KB) entities on the other, filter, rank and recommend citation-worthy documents. The pipeline encountered in systems that approach this problem involves four stages: filtering, classification, ranking (or scoring), and evaluation. Filtering is only an initial step that reduces the web-scale corpus into a working set of documents more manageable for the subsequent stages. Nevertheless, this step has a large impact on the recall that can be attained maximally. This study analyzes in-depth the main factors that affect recall in the filtering stage. We investigate the impact of choices for corpus cleansing, entity profile construction, entity type, document type, and relevance grade. Because failing on recall in this first step of the pipeline cannot be repaired later on, we identify and characterize the citation-worthy documents that do not pass the filtering stage by examining their contents.

This paper provides an overview of our participation in the Contextual Suggestion Track. The TREC 2014 Contextual Suggestion Track allowed participants to submit personalized rankings using documents either from the Open Web or from an archived, static Web collection (ClueWeb12) collection. One of the main steps in recommending attractions for a particular user in a given context is the selection of the candidate documents. This task is more challenging when relying on ClueWeb12 collection rather than public tourist APIs for finding suggestions. In this paper, we present our approach for selecting candi- date suggestions from the entire ClueWeb12 collection using the tourist domain knowledge available in the Open Web. We show that the generated recommendations to the provided user profiles and contexts improve significantly using this inferred domain knowledge.

Cumulative Citation Recommendation (CCR) is defined as: given a stream of documents on one hand and Knowledge Base (KB) entities on the other, filter, rank and recommend citation-worthy documents. The pipeline encountered in systems that approach this problem involves four stages: filtering, classification, ranking (or scoring), and evaluation. Filtering is only an initial step that reduces the web-scale corpus into a working set of documents more manageable for the subsequent stages. Nevertheless, this step has a large impact on the recall that can be at- tained maximally. This study analyzes in-depth the main factors that affect recall in the filtering stage. We investigate the impact of choices for corpus cleansing, entity profile construction, entity type, document type, and relevance grade. Because failing on recall in this first step of the pipeline cannot be repaired later on, we identify and characterize the citation-worthy documents that do not pass the filtering stage by examining their contents.

The Visual Dependency Representation (VDR) is an explicit model of the spatial relationships between objects in an image. In this paper we present an approach to training a VDR Parsing Model without the extensive human supervision used in previous work. Our approach is to find the objects mentioned in a given description using a state-of-the-art object detector, and to use successful detections to produce training data. The description of an unseen image is produced by first predicting its VDR over automatically detected objects, and then generating the text with a template-based generation model using the predicted VDR. The performance of our approach is comparable to a state-of-the-art multimodal deep neural network in images depicting actions.

We investigate the role of geographic proximity in news consumption. Using a month-long log of user interactions with news items of ten information portals, we study the relationship between users' geographic locations and the geographic foci of information portals and local news categories. We find that the location of news consumers correlates with the geographical information of the information portals at two levels: the portal and the local news category. At the portal level, traditional mainstream news portals have a more geographically focused readership than special interest portals, such as sports and technology. At a finer level, the mainstream news portals have local news sections that have even more geographically focused readerships.

Traditional batch evaluation metrics assume that user interaction with search results is limited to scanning down a ranked list. However, modern search interfaces come with additional elements supporting result list refinement (RLR) through facets and filters, making user search behavior increasingly dynamic. We develop an evaluation framework that takes a step beyond the interaction assumption of traditional evaluation metrics and allows for batch evaluation of systems with and without RLR elements. In our framework we model user interaction as switching between different sublists. This provides a measure of user effort based on the joint effect of user interaction with RLR elements and result quality. We validate our framework by conducting a user study and comparing model predictions with real user performance. Our model predictions show significant positive correlation with real user effort. Further, in contrast to traditional evaluation metrics, the predictions using our framework, of when users stand to benefit from RLR elements, reflect findings from our user study. Finally, we use the framework to investigate under what conditions systems with and without RLR elements are likely to be effective. We simulate varying conditions concerning ranking quality, users, task and interface properties demonstrating a cost-effective way to study whole system performance.

Internet users are turning more frequently to online news as a replacement for traditional media sources such as newspapers or television shows. Still, discovering news events online and follow- ing them as they develop can be a difficult task. In previous work, we presented a novel approach to extract sentences from an online stream of news articles that summarizes the most important news facts for a given ad-hoc information need, which compared to ex- isting systems obtained relatively high-precision results and a com- parable recall [9]. In this track, we experiment with this approach to improve the recall of retrieved results.

Cumulative Citation Recommendation (CCR) is defined as: given a stream of documents on one hand and Knowledge Base (KB) entities on the other, filter, rank and recommend citation-worthy documents. The pipeline encountered in systems that approach this problem involves four stages: filtering, classification, ranking (or scoring), and evaluation. Filtering is only an initial step that reduces the web-scale corpus into a working set of documents more manageable for the subsequent stages. Nevertheless, this step has a large impact on the recall that can be at- tained maximally. This study analyzes in-depth the main factors that affect recall in the filtering stage. We investigate the impact of choices for corpus cleansing, entity profile construction, entity type, document type, and relevance grade. Because failing on recall in this first step of the pipeline cannot be repaired later on, we identify and characterize the citation-worthy documents that do not pass the filtering stage by examining their contents.

Traditional batch evaluation metrics assume that user interaction with search results is limited to scanning down a ranked list. However, modern search interfaces come with additional elements supporting result list refinement (RLR) through facets and filters, making user search behavior increasingly dynamic. We develop an evaluation framework that takes a step beyond the interaction assumption of traditional evaluation metrics and allows for batch evaluation of systems with and without RLR elements. In our framework we model user interaction as switching between different sublists. This provides a measure of user effort based on the joint effect of user interaction with RLR elements and result quality. We validate our framework by conducting a user study and comparing model predictions with real user performance. Our model predictions show significant positive correlation with real user effort. Further, in contrast to traditional evaluation metrics, the predictions using our framework, of when users stand to benefit from RLR elements, reflect findings from our user study. Finally, we use the framework to investigate under what conditions systems with and without RLR elements are likely to be effective. We simulate varying conditions concerning ranking quality, users, task and interface properties demonstrating a cost-effective way to study whole system performance.

This paper provides an overview of our participation in the Contextual Suggestion Track. The TREC 2014 Contextual Suggestion Track allowed participants to submit personalized rankings using documents either from the Open Web or from an archived, static Web collection (ClueWeb12) collection. One of the main steps in recommending attractions for a particular user in a given context is the selection of the candidate documents. This task is more challenging when relying on ClueWeb12 collection rather than public tourist APIs for finding suggestions. In this paper, we present our approach for selecting candi- date suggestions from the entire ClueWeb12 collection using the tourist domain knowledge available in the Open Web. We show that the generated recommendations to the provided user profiles and contexts improve significantly using this inferred domain knowledge.

We investigate the role of geographic proximity in news consumption. Using a month-long log of user interactions with news items of ten information portals, we study the relationship between users' geographic locations and the geographic foci of information portals and local news categories. We find that the location of news consumers correlates with the geographical information of the information portals at two levels: the portal and the local news category. At the portal level, traditional mainstream news portals have a more geographically focused readership than special interest portals, such as sports and technology. At a finer level, the mainstream news portals have local news sections that have even more geographically focused readerships.

Web archives attempt to preserve the fast changing web, yet they will always be incomplete. Due to restrictions in crawling depth, crawling frequency, and restrictive selection policies, large parts of the Web are unarchived and, therefore, lost to posterity. In this paper, we propose an approach to uncover unarchived web pages and websites and to reconstruct different types of descriptions for these pages and sites, based on links and anchor text in the set of crawled pages. We experiment with this approach on the Dutch Web Archive and evaluate the usefulness of page and host-level representations of unarchived content. Our main findings are the following: First, the crawled web contains evidence of a remarkable number of unarchived pages and websites, potentially dramatically increasing the coverage of a Web archive. Second, the link and anchor text have a highly skewed distribution: popular pages such as home pages have more links pointing to them and more terms in the anchor text, but the richness tapers off quickly. Aggregating web page evidence to the host-level leads to significantly richer representations, but the distribution remains skewed. Third, the succinct representation is generally rich enough to uniquely identify pages on the unarchived web: in a known-item search setting we can retrieve unarchived web pages within the first ranks on average, with host-level representations leading to further improvement of the retrieval effectiveness for websites.

Cumulative Citation Recommendation (CCR) is defined as: given a stream of documents on one hand and Knowledge Base (KB) entities on the other, filter, rank and recommend citation-worthy documents. The pipeline encountered in systems that approach this problem involves four stages: filtering, classification, ranking (or scoring), and evaluation. Filtering is only an initial step that reduces the web-scale corpus into a working set of documents more manageable for the subsequent stages. Nevertheless, this step has a large impact on the recall that can be attained maximally. This study analyzes in-depth the main factors that affect recall in the filtering stage. We investigate the impact of choices for corpus cleansing, entity profile construction, entity type, document type, and relevance grade. Because failing on recall in this first step of the pipeline cannot be repaired later on, we identify and characterize the citation-worthy documents that do not pass the filtering stage by examining their contents.

The Visual Dependency Representation (VDR) is an explicit model of the spatial relationships between objects in an image. In this paper we present an approach to training a VDR Parsing Model without the extensive human supervision used in previous work. Our approach is to find the objects mentioned in a given description using a state-of-the-art object detector, and to use successful detections to produce training data. The description of an unseen image is produced by first predicting its VDR over automatically detected objects, and then generating the text with a template-based generation model using the predicted VDR. The performance of our approach is comparable to a state-of-the-art multimodal deep neural network in images depicting actions.

Tasks that require users to have expert knowledge are diffi- cult to crowdsource. They are mostly too complex to be carried out by non-experts and the available experts

Many national and international heritage institutes realize the importance of archiving the web for future culture heritage. Web archiving is currently performed either by harvesting a national domain, or by crawling a pre-defined list of websites selected by the archiving institution. In either method, crawling results in more information being harvested than just the websites intended for preservation; which could be used to reconstruct impressions of pages that existed on the live web of the crawl date, but would have been lost forever. We present a method to create representations of what we will refer to as a web collection's (aura): the web documents that were not included in the archived collection, but are known to have existed --- due to their mentions on pages that were included in the archived web collection. To create representations of these unarchived pages, we exploit the information about the unarchived URLs that can be derived from the crawls by combining crawl date distribution, anchor text and link structure. We illustrate empirically that the size of the aura can be substantial: in 2012, the Dutch Web archive contained 12.3M unique pages, while we uncover references to 11.9M additional (unarchived) pages.