Can We Find Documents in Web Archives without Knowing their Contents?

TL;DR

This paper explores the potential of using metadata evidence from web archive documents, such as headers and URLs, to improve search and ranking without relying on full content analysis.

Contribution

It introduces a novel approach that leverages metadata evidence for ranking web archive documents, demonstrating its effectiveness through empirical evaluation.

Findings

Metadata evidence can effectively distinguish relevant from irrelevant documents.

The proposed model outperforms content-based methods in certain scenarios.

Using metadata reduces computational costs and handles high redundancy.

Abstract

Recent advances of preservation technologies have led to an increasing number of Web archive systems and collections. These collections are valuable to explore the past of the Web, but their value can only be uncovered with effective access and exploration mechanisms. Ideal search and rank- ing methods must be robust to the high redundancy and the temporal noise of contents, as well as scalable to the huge amount of data archived. Despite several attempts in Web archive search, facilitating access to Web archive still remains a challenging problem. In this work, we conduct a first analysis on different ranking strategies that exploit evidences from metadata instead of the full content of documents. We perform a first study to compare the usefulness of non-content evidences to Web archive search, where the evidences are mined from the metadata of file headers, links and URL strings only. Based on these findings, we propose a simple yet surprisingly effective learning model that combines multiple evidences to distinguish "good" from "bad" search results. We conduct empirical experiments quantitatively as well as qualitatively to confirm the validity of our proposed method, as a first step towards better ranking in Web archives taking meta- data into account.

Tables

Table 1. Table 1 : List of link types and their patterns, as used to extract from the Internet Archive data

HTML Tag pattern	Destination type
A/href	Either a hyperlink or an anchor
IMG/src	Image source
AREA/href	Area inside an image-map
EMBED/src	Embeded object
FRAME/src	Frame source
INPUT/src	Input source
IFRAME/src	Iframe source
FORM/action	Form processing URL
TD/background	Table cell background image
TR/background	Table row background image
BODY/background	Page background image
OBJECT/codebase	Code of embeded object
TABLE/background	Table background image
FB:LOGIN-BUTTON/background	(Facebook login) button background image

Table 2. Table 2: List of Entity types

Type	Examples
Politician	Barack Obama, Angela Merkel
Scientist	Max Planck, Josef Meixner
Artist	Michael Jackson, Scorpions
Sport player	Andreas Nilsson, Franz Beckenbauer
Author	Franz Hoellering, Artur Kaps
Entrepreneur	Enzo Ferrari, Artur Mest
Organisation	Maharashtra Film Company
Product	Chevrolet Vega, Apple Macintosh
Location	Volken, Larrabee State Park
Works	Volken, Alien Trespass
Event	FIS-Ladies-Winter-Tournee 2011
Biology	Rote Mangrove, Ortolan
Music	Appenzeller Streichmusik
Astrology	CloudSat, Abell 262
Abstract Concept	Aequat Causa Effectum

Table 3. Table 3 : Selected features for analysing the ranking models, as grouped into different categories of evidence sources

Source	Name	Description
URL	URLDepth	The depth of the document URL
	QueryString	Whether the document URL contains query strings
	SearchWord	Appearance of searching words (“such”,“suchergebnis”, “query=”, etc.) in the document URL
	QueryInURL	Frequency of query terms in the document tokenized URL
	NewsURL	Whether the document URL belongs to a news domain
	WikipediaURL	Whether the URL has been cited in the Wikipedia page of the queried entity
Links	InlinkCount	The number of incoming links pointing to the document in the Web archive
	PageRankCore	PageRank score of the document in the hyperlink graph
	PageRankDomain	PageRank score of the domain in the hyperlink graph, where nodes are projected to domains
Anchor Texts	AnchorFreq	The fraction of incoming links pointing to the document in the Web archive, and having queries in their anchor texts
	AnchorTimeSpans	Time spans of two consecutive timestamps of the anchor texts that are longer than 1 week
	LuceneTf	Max value of Lucene term frequencies [10]
Metadata	DocLen	Document length counted by anchor texts
	FieldNorm	Lucene field-length normalization score of anchor texts [10]
	Idf	Inverse frequency of documents contain the query terms [10]
	Revision	Number of revisions of the document
	RevDuration	Count of time spans of two consecutive revisions that are longer than 1 week
	DomainSize	Count of web pages having the same domain as the document
	EntType	Type of entity represented by the query

Table 4. Table 4 : Retrieval performance using Bing ranks as ground truth (RF-B). Symbol ▲ indicates significant improvement over BM25.

	P@1	P@10	NDCG@10	MAP
BM25	0.153	0.123	0.527	0.173
PageRank	0.077	0.085	0.274	0.042
QueryInURL	0.154	0.185	0.700	0.202
RF	0.870▲	0.736▲	0.754▲	0.804▲

Table 5. Table 5 : Retrieval Performance based on manual labels (RF-M). Symbol ▲ indicates significant improvement over BM25.

	P@1	P@10	NDCG@10	MAP
BM25	0.385	0.269	0.517	0.270
PageRank	0.077	0.277	0.508	0.261
QueryInURL	0.384	0.353	0.553	0.332
RF	0.846▲	0.769▲	0.760▲	0.798▲

Table 6. Table 6 : Top Features by Information Gain

RF-M	RF-B
1. InlinkCount	1. Revision
2. QueryInURL	2. InlinkCount
3. Revision	3. DomainSize
4. Idf	4. QueryInURL
5. DomainSize	5. URLDepth
6. PageRankDomain	6. Idf
7. PageRankCore	7. AnchorTimeSpans
8. LuceneTf	8. FieldNorm
9. AnchorTimeSpans	9. LuceneTf
10. AnchorFreq	10. PageRankDomain
11. URLDepth	11. AnchorFreq

Table 7. Table 7 : Comparison of top- 5 5 5 results between RF models learnt using using manual labels and Bing rank. ∗ ∗ \ast indicates URLs that are not alive anymore in the Web, whereas ∗ ⁣ ∗ ∗ ∗ \ast\ast indicates URLs that are now redirected ones, checked in April 2016.

Query	RF-M	RF-B
Angela Merkel	wiwo.de/koepfe-der-wirtschaft/angela-merkel/5288044.html	spiegel.de/international/topic/angela_merkel
	spiegel.de/international/topic/angela_merkel	spiegel.de/thema/angela_merkel
	spiegel.de/thema/angela_merkel	sueddeutsche.de/thema/angela_merkel
	sueddeutsche.de/thema/angela_merkel	angela-merkel.de/
	zeit.de/schlagworte/personen/angela-merkel/index∗∗	welt.de/themen/angela-merkel/2
Albert Einstein	spiegel.de/thema/albert_einstein	spiegel.de/thema/albert_einstein
	amazon.de/albert-einstein-biographie-suhrkamp-taschenbuch/dp/3518389904	zitate-online.de/autor/einstein-albert
	zitate-online.de/autor/einstein-albert	einsteingalerie.de/
	sueddeutsche.de/thema/albert_einstein	einstein-website.de/z_kids/biographiekids.htm∗
	einstein-gymnasium-hameln.de/schule/einstein/einstein.php	helles-koepfchen.de/albert_einstein
Volkswagen	volkswagen.de/de.html	volkswagen.de/de.html
	autohaus24.de/neuwagen-kaufen/volkswagen	auto.de/kfzkatalog/vw
	volkswagen-nutzfahrzeuge.de/de.html	motor-talk.de/forum/volkswagen-b22.html
	motor-talk.de/forum/volkswagen-b22.html	volkswagen-nutzfahrzeuge.de/de.html
	spiegel.de/thema/vw	autohaus24.de/neuwagen-kaufen/volkswagen
Bruce Willis	sueddeutsche.de/thema/bruce_willis	sueddeutsche.de/thema/bruce_willis
	amazon.de/bruce-willis/e/b000apunwa∗	kino.de/star/bruce-willis/8453∗∗
	new-video.de/darsteller-bruce-willis	filmstarts.de/personen/197-bruce-willis.html
	kino.de/star/bruce-willis/8453∗∗	moviemaze.de/celebs/15/1.html∗∗
	moviemaze.de/celebs/15/1.html∗∗	new-video.de/darsteller-bruce-willis

Full text

In this work, we conduct a first analysis on different ranking strategies that exploit evidences from metadata instead of the full content of documents. We perform a first study to compare the usefulness of non-content evidences to Web archive search, where the evidences are mined from the metadata of file headers, links and URL strings only. Based on these findings, we propose a simple yet surprisingly effective learning model that combines multiple evidences to distinguish “good” from “bad” search results. We conduct empirical experiments quantitatively as well as qualitatively to confirm the validity of our proposed method, as a first step towards better ranking in Web archives taking metadata into account.

††conference: WebSci ’16, May 22-25, 2016, Hannover, Germany

{CCSXML}

<ccs2012> <concept> <concept_id>10002951.10003317.10003318</concept_id> <concept_desc>Information systems Document representation</concept_desc> <concept_significance>500</concept_significance> </concept> <concept> <concept_id>10002951.10003317.10003338.10003343</concept_id> <concept_desc>Information systems Learning to rank</concept_desc> <concept_significance>300</concept_significance> </concept> </ccs2012>

\ccsdesc

[500]Information systems Document representation \ccsdesc[300]Information systems Learning to rank

\printccsdesc

1 Introduction

The Web exhibits rapid growth and dynamic changes of its structure and contents, where ephemeral web pages are dominant, and content disappearance is commonplace [22]. A study by Toyoda et al. [24] suggests that more than half of the URLs disappeared or changed their location within one year. In order to preserve parts of the Web for future generations, many Web archiving initiatives are active, among which the Internet Archive is the most prominent one. Web archive data, which can be up to petabytes or more in size, consist of several snapshots of contents crawled from the Web at different points in time. Such collections are valuable for journalists, economists, historians, social scientists and others to explore the past, but their values can only be uncovered with an effective access and exploration mechanism. Such a mechanism should rank documents not only by their content relevance to an information need, as in the case of the actual Web, but also by their long-term preservation values reflected in the archive. Furthermore, Web archives are characterized by the high redundancy of contents, the imbalance of revisions due to selective crawling strategies [12], and thus call for different ranking approaches.

Searching in Web archives has drawn increasing attention in recent years. Early attempts focus on efficiently indexing the full contents of documents, with optimization tailored to special types of queries [3]. Other recent work suggests that non-content evidences such as timestamp [21], hyperlinks [7], or anchor texts [8, 12] can be used to improve ranking performance. For example, the anchor texts of hyperlinks, when aggregated by the target page, can well represent the collective perception of the document essence [6], in contrast to the actual content of the document which reflects the author’s intent. Similarly, authority evidence of a web page such as its PageRank score can complement the relevance of the page with its importance or freshness [21]. Despite several approaches having been proposed and developed, no optimal search strategy for Web archives has been developed yet. Furthermore, while the existing work suggests the benefit of different individual evidences, no study has compared the impacts of these evidences in a unified ranking framework, their contribution as well as disadvantages when put altogether.

In this paper, we follow the above school of thought, and argue that using other information than the document contents can help finding relevant and important documents with adequate performance, while avoiding the high cost of fully indexing the Web archive. Furthermore, we provide the first study to compare the usefulness of non-content evidences in web archive search, where the evidences are mined from the document metadata such as URL strings, file headers, and from the links such as anchor texts and the linking statistics. We conduct experiments with an entity search scenario in mind, i.e. finding documents related to an entity from the Web archive, and focus on entities with low ambiguity to reduce the effect of spurious results. In summary, we investigate the following research questions:

• •

RQ1: How useful are non-content features of documents for finding relevant documents in a Web archive?

• •

RQ2: Does the combination of multiple features improve the performance over individual ones?

We explore the above questions on subsets of the Internet Archive dataset, with our experiments focusing on the .de domain. We investigate the problem of exploration in the Web archive, where the queries are unambiguous entities identifiable from Wikipedia, and the documents are web pages with several revisions captured in the archive111In this paper, we will use “document” and “web page” interchangeably..

We examine several features derived from different sources of the documents, without processing the full text of documents. While the results confirm our assumption that non-content evidences are valuable resources and deserve more attention, they also give interesting insights into the influences of features in different settings. Based on the findings of this analysis, we propose a simple yet effective ranking model to combine multiple evidences for distinguishing “good” from “bad” search results in a Web archive, without knowing their full contents, and a scalable approach to obtain the training data for our learning models without much human effort.

The remainder of the paper is organised as follows. After the discussion of related work in Section 2, in Section 3 we describe the dataset used in our study, the characteristics of the dataset as well as some first exploratory analyses. Section 4 reports the analysis in Web archive search using non-content evidences. In Section 5, we discuss our ranking model based on non-content features, and show that it is indeed able to distinguish “good” from “bad” search results. Finally, in Section 6, we close with a discussion of future work and conclusions.

2 Related Work

Related work can be categorised into three groups. The first group consists of work on building infrastructure for indexing and accessing Web archives. The second group consists of work on exploiting non-content features in information retrieval. As there exists a vast amount of work in this area, we focus on existing work on retrieving documents using anchor texts, as this provides several interesting questions for our analysis. The third group of work consists of ranking approaches in Web archives.

2.1 Indexing and Searching Web Archives

Web Archives are used to preserve digital heritage and the web for future use [18]. One of the major issue in this context is how to provide infrastructures for indexing, searching, and access to these often huge archives [23]. This lead to a bulk of technical difficulties which have to be solved. Stack [23] proposed a modification of Nutch to enable web archive indexing and full-text search on larger-scale collections, which increase the capacity of Nutch from under 100M to higher amounts. Lin et al. [17] provide a scalable tool using Map-Reduce paradigm to parse data from and to the index.

2.2 Exploiting Anchor Texts for Retrieval

Anchor text of links in documents has been widely adopted as promising information in the context of Web search. A number of studies find that anchor text is useful for information retrieval when queries are navigational [6, 11], as well as for ad hoc information search [15]. In [6], Craswell et al. aggregated anchor texts for a target page and used them as surrogate documents for finding sites. In [11], Fujii et al. studied the effectiveness of content-based or anchor-based retrieval methods for different query types. Both [6, 11] concluded that incorporating anchor text can significantly improve the retrieval effectiveness for navigational queries. Koolen et al. [15] further investigated the importance of anchor text for ad hoc information search and showed that anchor text based methods can significantly improve retrieval effectiveness.

In the context of Web archive search, anchor texts can help in additional interesting ways. For example, they can be used to address the incompleteness of the archive collections, and help retrieving unarchived Web pages, or Web pages that were once present in the Web but not preserved in the archives. Klein and Nelson [13] leveraged the top-$n$ words of link anchors to extract lexical signatures of such missing webpages, which were then used to retrieve alternative URLs for these missing webpages. Similarly, Huurdeman et al. [12] conducted research to use links and anchors in the crawled webpages to recover unarchived webpages. Other research work combined anchor texts with additional information. Some researchers aggregated anchor texts from all pages that link to a page in the same domain [19], or the same website [14, 16] as the target page. Regarding time, historical trends of anchor texts have also been investigated for estimating anchor text importance. For instance, Dai et al. [8] differentiated pages with different in-link creation rates, and concluded that ranking performance can be improved via taking into account in-link creation rates. Nguyen et al. [20] study the problem of mining temporal subtopics in the Web Archive. They propose to mine the relevant time of temporal subtopics by leveraging the trending behaviour of the corresponding anchor texts.

In this paper, we focus on a different scenario, an exploratory search scenario, and find that anchor texts alone do not provide an optimal solution. We incorporate anchor text and other non-content features, without involving content of webpages, and show that this combination is better able to distinguish between good and bad search results.

2.3 Ranking Models in Web Archives

There is a number of approaches addressing the issue of ranking in the Web archives, which can be classified into three main lines. The first approach is based on full content indexing, suggested in the early work by Berberich et al. [2]. This approach takes into account temporal expressions and integrates them into a language model retrieval framework. However, the method only works for limited types of queries. The second approach consists of graph-based methods, exploiting the hyperlinks in the documents of the archive. In [7], Dai et al. leverage features from historical author activities and propose to consider the authority over multiple web snapshots at different time points. They modify the traditional link-based web ranking algorithms by further incorporating web page freshness over time from page and in-link activity. Nguyen et al. [21] attempt to discover content that can cover most interesting time periods for a given topic. To this end, they design a novel graph-based model by integrating relevance, temporal authority, diversity and time in a unified framework. A third line of research is represented by the learning approach in Costa et al. [5]. They assume that closer time periods are more likely to hold similar web characteristics. Based on this assumption, they propose a novel temporal-dependent ranking framework which exploits the variance of web characteristics over time. In contrast to this work, our work targets ranking documents and not revisions, and focuses on non-content features, to provide a more light-weight ranking framework.

3 Web Archive Data

3.1 Dataset

In this work, we use subsets of Web archive data provided by the Internet Archive from September 1996 to the end of December 2014. We focus our study on web pages with the .de top-level domains. This collection consists of 762,008 archive files, including 124,743 .arc files and 637,265 .warc files, which comprise 55.6TB in volume222All collection files are compressed in gzip format. It contains 1,434,118,956 URLs and 141,835,258,519 revisions, belonging to 9,880,121 of domains and 702,398,802 core URLs (i.e., URL obtained after excluding query strings). Revisions are not evenly distributed over URLs, but instead heavily biased to a small portion of domains, due to a number of selective crawling strategies [12]. This is known as the incompleteness problem, which is ubiquitous not only in the Internet Archive dataset, but in all Web archives [12]. We take this characteristic into account in our analysis in the next sections.

In this paper, we focus on investigating the usefulness of document metadata in Web archives without analysing the full contents. The metadata come from the following major sources: From the URL of the document, from the header of the archive files (.arc and .warc formats), and from the hyperlinks. As for the URL, we tokenize the string into different words, use a fixed set of defined delimiters (e.g. “/”, “.”, “-”, “_” ). The hyperlink information requires slightly more work to be extracted, we describe this separately in the following.

3.2 Preprocessing

Web Archive Graphs. Hyperlink information has been proven to be a powerful indicator for searching in Web archives [21, 7]. Inspired by previous work, we extracted the link structure of the Web archive documents in our set. There are 305,726,983,071 links between revisions, with 14 link types (shown in Table  1). Most of link types are links to background or images in documents. We select only links to content pages, identified by the tag <a>. There are 161,728,553,318 links after removing noisy pages. The existing links contain anchor texts to a document revision in the archive, with approximately 2 links per revision in average.

Web Archive Anchor Texts. One special aspect of a hyperlink is the anchor text, which is a visible text snippet used to label the links. To some extent, anchor texts encode the testimony of the users to the target, and when put together collectively, can represent a surrogate of the document. We extracted 161,728,553,318 distinct anchors text links from the hyperlinks, with 118 links per document on average.

Indexing. Inspired by early work on anchor text-based search [6], we represent a web document by concatenating all of their anchor texts pointing to this document, and utilize a full-text index to index this document “anchor representation”. We use ElasticSearch 333https://www.elastic.co/ to achieve indexing capacity at large scale. One subtle issue is that in Web archives, a link can be repeated multiple times due to crawling strategies. For such links, we decide to perform two strategies: (1) keep one unique anchor text for a source-destination pair for each revision, and (2) keep all anchor texts across revisions, even it is possible (and likely) that such revisions are identical. This allows us to exploit the accumulated statistics of anchor texts over time, and to analyse the effect of time on the anchoring behaviour444In strategy (2), we truncate the anchor texts of 26 URLs of which the length exceeds the storage limit of ElasticSearch..

There are 26,443,384,902 destination URLs for all links in web archive. However, we only make use of the ones which are archived in our dataset, resulting in 990,031,302 corresponding documents. Among these, there are 100,047,693 documents that have no anchor text at all, which we remove, reducing the number of retrievable documents to 889,983,609. This data is indexed by ElasticSearch as mentioned above. The resulting documents serve as potential search results from the archive.

3.3 Anchor Texts Distribution

As mentioned above, anchor texts have been widely used as a useful meta-data besides full contents to virtually represent the (target) document, with a rich body of related work (Section 2.2) In this paper, we investigate this further, by analysing the distribution of anchor texts in the context of Web archives.

Figures 1a-d show the distribution of anchor texts measured by the number of destination Web pages in log-log scales. From these figures, a heavy-tail distribution of anchor text usage can be observed. Another observation is that anchor texts follow power law distribution quite strictly in the head part, while the tail part exhibits high noise for popular anchor texts. The degree of noise increases over the years, as shown in the Figure 1d. In this figure, we group the distribution by projecting timestamps of source Web pages for each hyperlinks, grouping them on a yearly scale, drawing each distributions for each year, and putting them in one timeline from 2007 to 2013. From this timeline, we can see the continuous increase of noise in the tails over the years. We believe that this is mainly due to the imbalanced number of snapshots of websites according to the crawling strategies, which results in a bias in the sampling of domains in the archive (i.e., not all Web pages in the domain are crawled, or are crawled at different frequency).

To verify this assumption, in Figures 1a and 1b, we analyse and show the distributions of the top $350$ biggest domains, judged by the number of member Web pages preserved under the domain. For these domains, we observe that the crawling policy is more consistent over time, i.e. Web pages are all aggressively and comprehensively crawled regardless of their active time periods. This results in the balanced number of revisions crawled for both rare and popular anchor texts drawn from these domains. In the figures, the tail becomes much less noisy. This finding is interesting, as it reveals that the revision counts (driven by the crawling policies) have an implication on distribution of anchor texts, making anchor texts different from those normally drawn from the contents of the Web pages, suggesting that they should not be exploited in isolation, but rather in combination with other evidences such as the size and crawling frequency of domains. In the next sections, we look further into the correlation of features and the quality of retrieved documents.

4 Analysis on Web Archive Search

As we are interested in exploiting non-content information for search, we investigate the two research questions described in Section 1. For question 1, we focus on entity search scenarios, described below, as the anchor texts for entities are more intuitive and interpretable, and also less noisy. For question 2, we analyse the correlation between relevance of documents and different non-content features.

4.1 Entity Queries

Analysis Setup. In order to avoid spurious effects of ambiguous queries in Web archive search, we limit our queries to entities identifiable by a Wikipedia page. We choose Wikipedia pages that are not list pages, disambiguation pages, and contain no commas and round brackets which indicate potential ambiguity of entity names 555https://en.wikipedia.org/wiki/Wikipedia:Article_titles. Our query set consists of $216$ queries, chosen from entities with both high and low popularity666The full list of queries are omitted due to space limit, and available upon request from khoi@L3S.de. Specifically, we use the page view counts (extracted using Hedera [25]) of a Wikipedia page as proxy for the corresponding entity popularities, partition the view counts into three buckets (Top, Middle, Bottom), and randomly choose entities in each partition to build the sample. We also group entities into different types to facilitate high-level analysis. We first used DBpedia knowledge base [1] to resolve the types of the sampled entities, and manually mapped each entity into one most plausible, coarse-grained type. Table 2 lists the types studied in our work. We acknowledge that the list is still primitive, and leave a deeper analysis with a more advanced taxonomy for the future (for example, multiple classes per entity).

Example Analysis. Figure 2 shows the query “Angela Merkel” as covered in the Web (worldwide and in Germany), and as covered in our .de web archive. The Web coverage timelines are derived from Google Trend, while for the .de web archive, we derive them from our anchor text index described in Section 3. All values in the timelines are normalized to $[0,1]$ by dividing to the max value.

The first observation from these timelines is that query coverage on the Web follows bursty patterns, with big spikes reflecting important events, quite synchronized between German and international websites. For instance, the first spikes during the end of 2005 correspond to the event of Angela Merkel starting her duty as the chancellor of Germany. As for the archived anchor dataset, the trendings are only visible during the period of big events observed from the actual Web, i.e. the end of 2005 (her first term as the chancellor) and the period of 2011-2012 (active period of the Eurozone crisis). For many other medium size events, we do not observe peaks, and the ratio of coverage stays relatively stable. We believe that the main cause for this asynchrony stems from the fact that the purpose of putting an anchor text in a hyperlink from one content to another is to endorse the destination content, or to establish the contextual relevance between the two contents. This is different from queries on the actual Web, which directly encodes the user information need. The implication is that any existing work that relies purely on anchor text as the source of retrieval can only gain adequate performance on certain sets of queries such as finding specific URL [6], or searching for highly debated topics [21]. For other types of information need, combination with other evidences is necessary.

4.2 Document Metadata and Search

Analysis Setup. In this section, we investigate the influence of different non-content evidences on Web archive search. To start with, we choose the three main sources of evidences for the preliminary analysis:

1. URL: We choose to investigate the depth of the document URL, i.e., the level of sub-domains or directory in which the corresponding web page resides in the Web server. The lower URL depth corresponds to more general web pages such as home pages, while higher depth indicates that the document represented by the web page is likely to mention more specific information (such as product detail in an e-commerce site).

2. Anchor Text: We count the frequency of the query as appeared in the anchor text representation of the document (see Section 3, Indexing) as another evidence.

3. Revision Number: We count how many times a web page has been crawled. To some extent, this reflects the importance of the web page, as we observe popular and highly dynamic web pages to be captured more often than others.

Bing Search Engine as relevance proxy. For each entity query, we query our index to get all documents that contain the query either as part of anchor text, or in the URL string of the corresponding web page. We call this dataset $A$. This dataset contains $1,276,900$ URLs per query in average (median value is $37199$ URLs per query). We also issue the query to the Bing search engine and fetch the URLs from the top-$100$ results that overlap with our results from the index. We call this dataset $B$. The intuition for this setting is that we assume top results returned by Bing are more likely to be of higher quality and relevance to the query of interest. To somewhat reduce the bias introduced by time, we issued the query to Bing at different points in time over the last half year, and merge the top-100 URLs of all results.

Analysis. Figure 3 shows the distributions of the averaged evidence scores per query, as derived from result set $A$ and $B$. It is calculated by obtaining retrieved documents for each entity query (either returned by Bing or by the local ElasticSearch index), calculating the corresponding feature values for each document and then getting the average. We observe that documents returned on the Bing first page tend to have lower depth of URL strings as compared to documents in general (median $1.82$ compared to $2.87$), as well as higher number of revisions ($3.80$ compared to $2.74$), and appear in more links where the queries are used as the anchor texts ($10.00$ compared to $1.93$). On the other hand, these evidence scores of top results exhibit higher degree of diversity, reflected by the wider quartile windows in all three (for the Anchor Text evidence, the first and third quartiles of top result are $[4.56,16.81]$ compared to $[0.00,5.15]$ for all results).

In Figure 4, we further investigate the correlation of this evidence diversity and the entity types. The scores, which are obtained in the same way as for Figure 3, are presented in log scales, thus include some minus values (for instance, the query of “Abstract Concept” entity type appear only $0.59$ times on average in the top URLs returned by Bing). For different entity types, we observe a big difference in the scores. For example, queries about politicians or sport players tend to have results with deep URL strings, as they often appear in news or other well-organised websites. They are also captured more often than others, reflected in higher average revision counts. On the other hand, queries about authors (play writers, novelists) have the least crawled results, mainly because many of them have static contents. As for the URL depth and anchor frequency, the highest average scores are for queries about location. Analysing deeper, we observe that the location terms appear much more frequently in the result web pages, including boilerplated parts of the pages such as disclaimers, ads, headers, etc. This preliminary analysis suggests that in general, any search system that deals with entity queries must take into account different dimensions, for example the inherent semantics of the entity types.

5 Ranking by Multiple Evidences

Based on our previous analysis, in this section, we investigate the effectiveness of combining multiple evidences from metadata to distinguish good from bad search results from the Web archive. We extract different features and rely on learning-to-rank models to provide a unified ranking score.

5.1 Features

Table 3 lists the set of features investigated in our ranking models. In the following we explain some selected features.

NewsURL: This feature estimates the newsworthy of a content via its domain. We curate from the Wikipedia page “Liste_deutscher_Zeitungen” and from the web page “Paper Boy”777http://www.thepaperboy.com/germany/newspapers/country.cfm, among other sources, to obtain the list of 507 German news sites.

WikipediaURL: This feature indicates the credibility of the content via its domain. We assume that domains that have more citations from the Wikipedia page of the entity query are more topically relevant. For instance, many citations from the Wikipedia page of Michael Jackson are from entertainment websites, while in Max Planck’s Wikipedia page they are mainly from scientific websites. We measure this feature by the number of times a domain is cited in the Wikipedia page of the entity.

AnchorTimeSpans: Besides the anchor text frequency feature (mentioned in Section 4), we also consider the temporal dimension of the links. The intuition is that in the Web archive, the time when the anchor texts appear also reveals some information about how important the target is. If two anchor texts come from two links of revisions that have close timestamps, they might not count as completely separate endorsement. For example, if a Web page is crawled two times within a few hours, the contents might be just identical, and therefore counting the anchor texts two times will give false high endorsements to the target documents. Hence, for this feature, we measure the distance between any two consecutive timestamps of the anchor texts, and count the number of time spans of which the distance is longer than $1$ week.

RevDuration: This feature relies on the frequency of the crawlers to estimate the quality of the documents. If the documents have many revisions that are crawled within a short time period (e.g., a few days), it might have less authority than other documents with revisions crawled in different years (the Web page are still relevant to users, or are still referred to from other Web pages posted later). For this feature, we measure the number of durations between two consecutive revisions that are at least $1$ week long.

PageRank: We also calculate the authority scores of the document by constructing the graphs of different links. The PageRankCore scores are calculated on the graph of direct hyperlinks between two web pages, and the PageRankDomain scores are calculated based on links between domains.

5.2 Experiment Setup

5.2.1 Training Data and Sampling

To study how the features can be combined in a unified learning-to-rank model, we need the relevance feedback for documents in the Web archive, or the training data labels. Since there is no standard benchmark for this problem, and since the focus of our work is to study the influence of features coming from different evidences, in this first study, we choose two pragmatic approaches for building the training data labels, described subsequently. First, we choose a subset of 15 entity queries from the testbed for our further experiments and case study analysis. As judging all URLs returned in the dataset $A$ is infeasible, we decide to evaluate on a small random sample of the results, constructed as follows. For each feature dimension, we divide the results into three partitions according to the order of their normalized scores. Then we randomly pick $20$ to $50$ Web pages from each partition. The sample is then obtained by pooling all Web pages from all features. Our sampling algorithm makes sure there are overlaps between the samples of each feature pair, so as to reduce the total size of the pool. In practice, this results in $400$ sampled documents per query in average. Then, two labeling strategies are applied:

Soft Labeling: Similarly as the approach discussed in Section 4, we compare the results retrieved by our index with the top-$100$ results returned by Bing. If found, the label of the document is measured by the inverse of Bing rank, otherwise the document is labeled zero. This enables us to exploit the order in Bing search results for pairwise comparison in a learning-to-rank model. The advantage of this approach is that we can easily scale up training data construction in future work.

Manual Labeling: Besides the automated approach using Bing search results, we also experimented with manual labeling. Four evaluators annotated the documents by checking their content in the browser and in the Internet Archive archive.org site. Each evaluator annotated the documents from the sample on a three value scale: 0 means irrelevant, 1 means relevant but not important, and 2 means relevant and important documents. Here the notion of “importance” is defined by guiding the evaluator to think of the document in a long term exploration scenario, and estimate, whether the document contains lasting information about the queried entity or not. For example, a biography page about Albert Einstein should be labeled as 2, while an ad page about a product with Albert Einstein portrait should be labeled as 1. The average Cohen’s Kappa for the evaluators’ pairwise inter-agreement is $\kappa=0.75$, suggesting that the assessment task has fair cognitive complexity.

Adding Positive Examples: To make sure the training data have enough positive and negative labeled URLs of both strategies, after pooling Web pages by features, we include all results of our dataset $B$ into the sample. We acknowledge that the sample has a strong effect on the results and models learnt, and plan future experiments with more advanced sampling strategies and larger samples. However, our current samples at least allows us to answer the question how well we can distinguish between good and bad search results, even if we cannot extrapolate these results to our full dataset A, which contains many more potential search results.

5.2.2 Ranking Methods

Baselines. We consider the following baselines: 1) The BM25-based ranking score as returned from our anchor text-based index; 2) PageRank score; 3) Scores based on frequency of queries in the document URL string (QueryInURL). Each of these baselines corresponds to one source of evidence, discussed in the analyses in Section 4.

Learning Models. For the learning-to-rank method, we employ Random Forests (RF)[4] as implemented in the RankLib toolkit[9]. We use $5$-fold cross validation and Normalized Discounted Cumulative Gain (NDCG) scores at top-$10$ results to optimize the parameters. RF models are trained on both training data with soft labels derived from Bing (denoted RF-B) and with manual labels (denoted RF-M). In each case, we use the labels of the same strategy as ground truth for the evaluation (so RF-B is evaluated against Bing results and RF-M against human feedback). As we investigate in this first study the potential of using non-content features to retrieve documents, we leave sophisticated settings of cross-model and cross-label evaluations for the future.

5.3 Empirical Results

We use standard information retrieval metrics Precision (at different cut-off threshold of 1 and 10), NDCG as well as MAP for the evaluation.

Table 4 shows the performance of the baselines and of our learning model RF-B, using Bing ranks as soft labels. Table 5 shows the performance of the baselines and of our learning model RF-M, using manual labels. In both settings, PageRank performs poorly, mainly because it does not target directly the query-driven relevance of the documents. BM25 has better performance, and works better when judged by the manual labels than by comparing with Bing ranks. Given that this baseline is the same in both experiments, it reveals that our human relevance feedback is more relaxed compared to the Bing soft labels. A surprising result is that its performance is comparable to QueryInURL, and actually slightly worse when we expand the cut-off threshold (from top-$1$ to top-$10$ results). In all three metrics Precision, NDCG and MAP. In other words, for general entity search scenario with no special information need, it might be more useful to look into the URL string and guess whether a web page is a good resource to explore, than to examine the raw anchor texts without any other context.

Applying learning models introduces a significant improvement on the retrieval in both settings (both with $p$-value $<0.001$). Note however, that RF-M and RF-B values are not directly comparable, as they are evaluated against different ground truths. In addition, the high values are caused by our small samples which include many positive search results, so cannot be generalized to performance of our ranking model on our larger dataset A. We will investigate in future work how the two models perform on substantially larger scale datasets.

We also see that the NDCG scores of both models are lower than Precision at the same cut-off threshold. This suggests that although we are able to distinguish relevant and important documents from less relevant ones, the ranks in the top results are still not good as they should be (i.e. relevant but not important), leaving additional room for future improvement of the ranking models.

5.4 Feature Analysis

In Table 6, we show the top features ordered by their importance, as measured using Information Gain, of the two learning models. InlinkCount and Revision are among the most discriminating features. QueryInURL is also a very useful feature, ranked second in RF-M and 4th in RF-B. In our experiments, anchor text did not get into the top ranks, but stays at position 10th and 11th for the two models. If we incorporate this feature with time dimension, it becomes more discriminating (position 9 in RF-M and 7 in RF-B). In addition, in our RF-B model (i.e. for distinguishing Bing search results against others), except InLinkCount, all top-5 features are “light-weight” features such as Revision, QueryInURL, URLDepth which can be extracted easily from document metadata. For the DomainSize feature, we can also perform a one-time process job to compute the values. This will help us to scale to larger subsets of Web archives.

5.5 Case Study Analysis

To look at our results in more detail, in Table 7 we show the comparison of top-$5$ results as retrieved by the two models. To guarantee a fair comparison, we only retrieve the documents that are available in the results of both models, and show them in the same order as they are ranked in each result list.

Angela Merkel: The top-$1$ result using the model learnt by manual labels is the biography web page of Angela Merkel hosted by wiwo.de, not included at top rank for the Bing experiment. The pages at the 5th rank of both models are news topic pages, although the first one (zeit.de page) is no longer available in the actual Web, and is replaced by another page on the same topic.

Albert Einstein: The first-ranked document is the same for both models. For the other 4 top results, however, the model learnt by manual labels tends to be more diverse, with one result (einstein-gymnasium-hameln.de) having nothing to do with Albert Einstein as the physicist. In constrast, the model using Bing ranks gives better results, all biographic information to the physicist Albert Einstein. Also, it manages to find the relevant page einstein-website.de/z_kids/biographiekids.htm in the archive, which is no longer available on the Web. Looking into the user evaluation feedback, we see that users tend not to agree on marginal cases where Web pages about Albert Einstein actually discuss some distantly related information such as the places named after him, or Web sites simply using the name Albert Einstein to symbolize outstanding intelligence, so our manual labels are not really very reliable in this case.

Volkswagen: Both models give good results. Although the ranks of documents differ, it is difficult to judge which one is really better than the other. We believe the reason is that the underlying information need is very diverse: User can search for the company, or search for the car products, either for commercial purposes or to look up technical instructions. This ambiguity makes it difficult for both models to learn a sensible meaning from simple labeling approaches.

Bruce Willis: Both models agree on the first position, but RF-M gives second position to an Amazon page with a list of Bruce Willis’ films, which is now only available in archive.org. It also pushes redirect URLs (fourth and fifth positions), because of the anchor texts pointing to. Using Bing rank, RF-B delivers better quality, with all top 5 results biographic and topic pages about the movie star. Looking closely to the results, we can see that most of the top results are about movies where Bruce Willis acted and not about his information as a celebrity. Perhaps such information need (search for Bruce Willis’ films) is more popular, as Bruce Willis is not known for many scandals and thus solicits less personal comments on the Web than other celebrities. This bias is handled better by relying on Bing labeling, rather than a small number of users labeling results.

5.6 Discussion

In general, for top results, both models agree in many cases, with some small difference between queries such as “Angela Merkel”. Comparing the analysis in Section 5.5 with our empirical evaluation in Section 5.3, we see that in several cases the model learnt using Bing ranks as soft labels tends to prefer documents which are more focused on the entity and also more important. As the results of Bing come from advanced models with a much richer set of dimensions and features, the integration of more features into our model will certainly enable it to reflect better the main aspects embedded in current search engine rankings. Given the small size of samples in our evaluation and the inclusion of all “good” results (from Bing), it would be interesting to see how good our models can deal with larger subsets of the archive, which will be more noisy and include more spamming page.

For entity queries, we see that some highly valuable pages, such as Wikipedia pages for the biographical information of the entities, are not shown in the top results. It is interesting to see how we can encode such credits with more features, tailored to entity exploration scenarios. In this context, other issues such as spamming URLs, user qualitative evaluation, etc. should be analysed deeper. For instance, the comparison between two models should not be limited to quantitative assessment, but also be extended to the user satisfaction measurement, to really understand the usefulness of the document non-content evidences. We can investigate further the search scenarios for different types of entities (e.g., the information need when user searches for a celebrity is substantially different than when she searches for a scientist, or a travelling location), and design different per-type questions for user experiments.

6 Conclusion

In this paper, we conducted a first study about the influence of non-content evidences on searching in Web archives. By using Bing search results as proxy for relevance, and by focusing our search scenario on exploring unambiguous entities, we can identify the correlation between different features and the quality of documents. We find out that although anchor texts are useful resources for searching in the Web, applying them to Web archives require more thoughts, taking into account other evidences such as entity types, URL, etc.

We also study the possibility of combing multiple evidences to improve the ranking. The result is promising, and can be a good starting point for an efficient ranking framework that does not rely on full content indexing of documents, at least for an entity exploration scenario, where the user has no special ad-hoc information need.

There are several other directions to extend this work. For instance, we aim to target different search scenarios in the Web archive, such as navigational search for particular (but partially forgotten) resources in the past. We also aim to improve the retrieval component of our work, to facilitate both ambiguous and unambiguous queries, and to scale and evaluate our results for larger subsets of Web archives.

Acknowledgements. This work was funded by the ERC Advanced Grant ALEXANDRIA (under the grant No. 339233). We thank the reviewers for the fruitful discussion and suggestions on future work. We also thank our colleague Philipp Kemkes for setting up the usage of Bing search API to facilitate our experiments.