The goal of the INEX 2013 Social Book Search Track is to evaluate approaches for supporting users in reading, searching, and navigating collections of books based on book metadata, the full texts of digitised books or associated user-generated content. The investigation is focused around three tasks: 1) the Social Book Search (SBS) task investigates the complex nature of relevance in book search and the role of traditional and user-generated book metadata in retrieval, 2) the Prove It (PI) task evaluates focused retrieval approaches for searching pages in books that can con rm or refute a given factual claim, 3) the Structure Extraction (SE) task evaluates automatic techniques for deriving book structure from OCR text and layout information. Both the SBS and SE tasks have a growing number of active participants, while the PI task is only tackled by a small number of core groups. In the SBS task, we extended last year's investigation into the nature of book suggestions from the LibraryThing forums and how they compare to book relevance judgements. We found further support that such suggestions are a valuable alternative to traditional test collections that are based on top-k pooling and editorial relevance judgements. The PI task added a further relevance criterion that pages should not only con rm or refute a given factual claim, but should also come from an authoritative source that is of the appropriate genre. The relevance assessments have not yet commenced at the time of writing. The SE task has reached a record number of active participants and has, for the rst time, witnessed an improvement in the state of the art.
For centuries books were the dominant source of information, but how we acquire, share, and publish information is changing in fundamental ways due to the Web. The goal of the Social Book Search Track is to investigate techniques to support users in searching and navigating the full texts of digitised books and complementary social media as well as providing a forum for the exchange of research ideas and contributions. Towards this goal the track is building appropriate evaluation benchmarks, complete with test collections for social, semantic and focused search tasks. The track provides opportunities to explore research questions around four key areas: { Evaluation methodologies for book search tasks that combine aspects of retrieval and recommendation, { Information retrieval techniques for dealing with professional and user-generated metadata, { Semantic and focused retrieval techniques for searching collections of digitised books, and { Mechanisms to increase accessibility to the contents of digitised books.
Based around these main themes, the following three tasks were de ned: 1. The Social Book Search (SBS) task, framed within the scenario of a user searching a large online book catalogue for a given topic of interest, aims at exploring techniques to deal with complex information needs|that go beyond topical relevance and can include aspects such as genre, recency, engagement, interestingness, and quality of writing|and complex information sources that include user pro les, personal catalogues, and book descriptions containing both professional metadata and user-generated content. 2. The Prove It (PI) task aims to evaluate focused retrieval approaches on collections of digitised books, where users expect to be pointed directly at relevant pages that may help to con rm or refute a given factual claim; 3. The Structure Extraction (SE) task aims at evaluating automatic techniques for deriving structure from ocr and building hyperlinked table of contents.
In this paper, we report on the setup and the results of each of these tasks at the 2013 inex@clef Lab. First, in Section 2, we give a brief summary of the participating organisations. The SBS task is described in detail in Section 3, the PI task in Section 4, and the SE task in Section 5. We close in Section 6 with a summary and plans for 2014. 2
A total of 68 organisations registered for the track (compared with 55 in 2012 and 47 in 2011). At the time of writing, we counted 14 active groups (compared with 5 in 2012 and 10 in 2011), see Table 1. 3
The goal of the Social Book Search (SBS) task is to evaluate the value of professional metadata and user-generated content for book search on the Web and 4 University of Amsterdam, ILLC SBS, PI 4 University of Amsterdam, ILPS SBS 54 Royal School of Library and Information Science SBS 100 Oslo & Akershus University College of Applied Sciences SBS, PI 113 University of Caen SE 123 LSIS / Aix-Marseille University SBS 147 National Taiwan Normal University SBS 180 Chaoyang University of Technology SBS 232 Indian School of Mines, Dhanbad SBS 280 University of Wurzburg SE 288 University of Innsbruck SE 299 EPITA/LRDE SE 303 Nankai University SE
Microsoft Development Center Serbia SE
Runs 6 SBS, 5 PI 2 SBS 3 SBS 3 SBS, 6 PI
5 SE 3 SBS 6 SBS 6 SBS 5 SBS 1 SE 1 SE 1 SE 1 SE 1 SE Total 34 SBS, 11 PI, 10 SE to develop and evaluate systems that can deal with both retrieval and recommendation aspects, where the user has a speci c information need against a background of personal tastes, interests and previously seen books. Through social media, book descriptions have extended far beyond what is traditionally stored in professional catalogues. Not only are books described in the users' own vocabulary, but are also reviewed and discussed online, and added to online personal catalogues of individual readers. This additional information is subjective and personal, and opens up opportunities to aid users in searching for books in di erent ways that go beyond the traditional editorial metadata based search scenarios, such as known-item and subject search. For example, readers use many more aspects of books to help them decide which book to read next [7], such as how engaging, fun, educational or well-written a book is. In addition, readers leave a trail of rich information about themselves in the form online pro les which contain personal catalogues of the books they have read or want to read, personally assigned tags and ratings for those books and social network connections to other readers. This results in a search task that may require a di erent model than traditional ad hoc search [5] or recommendation.
The SBS task investigates book requests and suggestions from the LibraryThing (LT) discussion forums as a way to model book search in a social environment. The discussions in these forums show that readers frequently turn to others to get recommendations and tap into the collective knowledge of a group of readers interested in the same topic.
The task builds on the INEX Amazon/LibraryThing (A/LT) collection [
The SBS task aims to address the following research questions: { Can we build reliable and reusable test collections for social book search based on book requests and suggestions from the LT discussion forums? { Can user pro les provide a good source of information to capture personal, a ective aspects of book search information needs? { How can systems incorporate both speci c information needs and general user pro les to combine the retrieval and recommendation aspects of social book search? { What is the relative value of social and controlled book metadata for book search? 3.1
The scenario is that of a user turning to Amazon Books and LT to nd books to read, to buy or to add to their personal catalogue. Both services host large collaborative book catalogues that may be used to locate books of interest.
On LT, users can catalogue the books they read, manually index them by assigning tags, and write reviews for others to read. Users can also post messages on discussion forums asking for help in nding new, fun, interesting, or relevant books to read. The forums allow users to tap into the collective bibliographic knowledge of hundreds of thousands of book enthusiasts. On Amazon, users can read and write book reviews and browse to similar books based on links such as \customers who bought this book also bought... ".
Users can search online book collections with di erent intentions. They can search for speci c known books with the intention to obtain them (buy, download, print). Such needs are addressed by standard book search services as o ered by Amazon, LT and other online bookshops as well as traditional libraries. In other cases, users search for a speci c, but unknown, book with the intention of identifying it. Another possibility is that users are not looking for a speci c book, but hope to discover one or more books meeting some criteria. These criteria can be related to subject, author, genre, edition, work, series or some other aspect, but also more serendipitously, such as books that merely look interesting or fun to read or that are similar to a previously read book. 3.2
The SBS task is to reply to a user request posted on a LT forum (see Section 3.5) by returning a list of recommended books matching the user's information need. More speci cally, the task assumes a user who issues a query to a retrieval system, which then returns a (ranked) list of relevant book records. The user is assumed to inspect the results list starting from the top, working down the list until the information need has been satis ed or until the user gives up. The retrieval system is expected to order the search results by relevance to the user's information need.
The user's query can be a number of keywords, but also one or more book records as positive or negative examples. In addition, the user has a personal pro le that may contain information on the user's interests, list of read books and connections with other readers. User requests may vary from asking for books on a particular genre, looking for books on a particular topic or period or books written in a certain style. The level of detail also varies, from a brief statement to detailed descriptions of what the user is looking for. Some requests include examples of the kinds of books that are sought by the user, asking for similar books. Other requests list examples of known books that are related to the topic but are speci cally of no interest. The challenge is to develop a retrieval method that can cope with such diverse requests.
The books must be selected from a corpus that consists of a collection of curated and social book metadata, extracted from Amazon Books and LT, extended with associated records from library catalogues of the Library of Congress and the British Library (see the next section). Participants of the SBS task are provided with a set of book search requests and user pro les and are asked to submit the results returned by their systems as ranked lists.
The SBS task, thus, combines aspects from retrieval and recommendation. On the one hand, the task is akin to directed search, familiar from information retrieval, with the requirement that returned books should be topically relevant to the user's information need described in the forum thread. On the other hand, users may have particular preferences for writing style, reading level, knowledge level, novelty, unusualness, presence of humorous elements and possibly many other aspects. These preferences are to some extent re ected by the user's reading pro le represented by the user's personal catalogue. This catalogue contains the books already read or earmarked for future reading, and may contain personally assigned tags and ratings. Such preferences and pro les are typical in recommendation tasks, where the user has no speci c information need, but is looking for suggestions of new items based on previous preferences and history. 3.3
We want to evaluate the book ranking of retrieval systems, speci cally the top ranks. We adopt the submission format of TREC, with a separate line for each retrieval result, consisting of six columns: 1. topic id: the topic number, which is based on the LT forum thread number. 2. Q0: the query number. Unused, so should always be Q0. 3. isbn: the ISBN of the book, which corresponds to the le name of the book description. 4. rank: the rank at which the document is retrieved. 5. rsv: retrieval status value, in the form of a score. For evaluation, results are ordered by descending score. 6. run id: a code to identify the participating group and the run.
Participants are allowed to submit up to six runs, of which at least one should use only the title eld of the topic statements (the topic format is described in Section 3.5). For the other ve runs, participants could use any eld in the topic statement. 3.4
To study the relative value of social and controlled metadata for book search,
we need a large collection of book records that contains controlled subject
headings and classi cation codes as well as social descriptions such as tags and
reviews, for a set of books that is representative of what readers are searching for.
We use the Amazon/LibraryThing (A/LT) corpus crawled by the University of
Duisburg-Essen for the INEX Interactive Track [
The collection consists of 2.8 million book records from Amazon, extended with social metadata from LT. This set represents the books available through Amazon. The records contain title information as well as a Dewey Decimal Classi cation (DDC) code (for 61% of the books) and category and subject information supplied by Amazon. We note that for a sample of Amazon records we noticed the subject descriptors to be noisy, with a number of inappropriately assigned descriptors that seem unrelated to the books.
Each book is identi ed by an ISBN. Since di erent editions of the same work have di erent ISBNs, there can be multiple records for a single intellectual work. Each book record is an XML le with elds like isbn, title, author, publisher, dimensions, numberofpages and publicationdate. Curated metadata comes in the form of a Dewey Decimal Classi cation in the dewey eld, Amazon subject headings in the subject eld, and Amazon category labels in the browseNode elds. The social metadata from Amazon and LT is stored in the tag, rating, and review elds. The full list of elds is shown in Table 2.
To ensure that there is enough high-quality metadata from traditional library catalogues, we extended the A/LT data set with library catalogue records from the Library of Congress (LoC) and the British Library (BL). We only use library records of ISBNs that are already in the A/LT collection. These records contain formal metadata such as title information (book title, author, publisher, etc.), classi cation codes (mainly DDC and LCC) and rich subject headings based on the Library of Congress Subject Headings (LCSH).5 Both the LoC records and the BL records are in MARCXML6 format. There are 1,248,816 records from the LoC and 1,158,070 records in MARC format from the BL. Combined, there are 2,406,886 records covering 1,823,998 of the ISBNs in the A/LT collection (66%). Although there is no single library catalogue that covers all books available on 5 For more information see: http://www.loc.gov/aba/cataloging/subject/ 6 MARCXML is an XML version of the well-known MARC format. See: http://www. loc.gov/standards/marcxml/ book similarproducts title imagecategory dimensions tags edition name reviews isbn dewey role editorialreviews ean creator blurber images binding review dedication creators label rating epigraph blurbers listprice authorid rstwordsitem dedications manufacturer totalvotes lastwordsitem epigraphs numberofpages helpfulvotes quotation
rstwords publisher date seriesitem lastwords height summary award quotations width editorialreview browseNode series length content character awards weight source place browseNodes readinglevel image subject characters releasedate imageCategories similarproduct places publicationdate url tag subjects studio data Amazon, we reason that these combined library catalogues can improve both the quality and quantity of professional book metadata. Indeed, with the LoC and BL data sets combined, 79% of all ISBNs in the original A/LT corpus now have a DDC code. In addition, the LoC data set also has LCC codes for 44% of the records in the collection. With only the A/LT data, 57% of the book descriptions have at least one subject heading, but with the BL and LoC data added, this increases to 80%. Furthermore, the A/LT data often has only a single subject heading per book, whereas in the BL and LoC data sets, book descriptions typically have 2{4 headings (average 2.96). Thus, the BL and LoC data sets increase the coverage of curated metadata, such that the vast majority of descriptions in our data set include professionally assigned classi cation codes and subject headings.
ISBNs and intellectual works Each record in the collection corresponds to an ISBN, and each ISBN corresponds to a particular intellectual work. An intellectual work can have di erent editions, each with their own ISBN. The ISBN-to-work relation is a many-to-one relation. In many cases, we assume the user is not interested in all the di erent editions, but in di erent intellectual works. For evaluation we collapse multiple ISBN to a single work. The highest ranked ISBN is evaluated and all lower ranked ISBNs of the same work ignored. Although some of the topics on LibraryThing are requests to recommend a particular edition of a work|in which case the distinction between di erent ISBNs for the same work are important|we ignore these distinctions to make evaluation easier. This turns edition-related topics into known-item topics.
However, one problem remains. Mapping ISBNs of di erent editions to a single work is not trivial. Di erent editions may have di erent titles and even have di erent authors (some editions have a foreword by another author, or a translator, while others have not), so detecting which ISBNs actually represent the same work is a challenge. We solve this problem by using mappings made by the collective work of LibraryThing members. LT members can indicate that two books with di erent ISBNs are actually di erent manifestations of the same intellectual work. Each intellectual work on LibraryThing has a unique work ID, and the mappings from ISBNs to work IDs is made available by LibraryThing.7
The mappings are not complete and might contain errors. Furthermore, the mappings form a many-to-many relationship, as two people with the same edition of a book might independently create a new book page, each with a unique work ID. It takes time for members to discover such cases and merge the two work IDs, which means that at any time, some ISBNs map to multiple work IDs even though they represent the same intellectual work. LibraryThing can detect such cases but, to avoid making mistakes, leaves it to members to merge them. The fraction of works with multiple ISBNs is small so we expect this problem to have a negligible impact on evaluation. 3.5
LT users discuss their books on the discussion forums. Many of the topic threads are started with a request from a member for interesting, fun new books to read. Users typically describe what they are looking for, give examples of what they like and do not like, indicate which books they already know and ask other members for recommendations. Members often reply with links to works catalogued on LT, which have direct links to the corresponding records on Amazon. These requests for recommendations are natural expressions of information needs for a large collection of online book records. We use a sample of these forum topics to evaluate systems participating in the SBS task.
Each topic has a title and is associated with a group on the discussion forums. For instance, topic 99309 in Figure 1 has the title Politics of Multiculturalism Recommendations? and was posted in the group Political Philosophy. The books suggested by members in the thread are collected in a list on the side of the topic thread (see Figure 1). A feature called touchstone can be used by members to easily identify books they mention in the topic thread, giving other readers of the thread direct access to a book record in LT, with associated ISBNs and links to Amazon. We use these suggested books as initial relevance judgements for evaluation. In the rest of this paper, we use the term suggestion to refer to a book that has been identi ed in a touchstone list for a given forum topic. Since all suggestions are made by forum members, we assume they are 7 See: http://www.librarything.com/feeds/thingISBN.xml.gz valuable judgements on the relevance of books. Additional relevance information can be gleaned from the discussions on the threads. Consider, for example, topic 1299398. The topic starter rst explains what sort of books he is looking for, and which relevant books he has already read or is reading. Other members post responses with book suggestions. The topic starter posts a reply describing which suggestions he likes and which books he has ordered and plans to read. Later on, after some more discussions, the topic starter provides feedback on the suggested books that he has now read. Such feedback can be used to estimate the relevance of a suggestion to the user.
User pro les and personal catalogues From LT we can not only extract the information needs of social book search topics, but also the rich user pro les of the topic starters, which contain information on which books they have in their personal catalogue on LT, which ratings and tags they assigned to them and a social network of friendship relations, interesting library relations and group memberships. These pro les may provide important signals on the user's topical and genre interests, reading level, which books they already know and which ones they like and don't like. These pro les were scraped from the LT site, transformed into xml and made available to participants. This adds a recommendation aspect to the task. One of the research questions of the SBS task is whether this pro le information can help systems in identifying good suggestions.
Although the user expresses her information need in some detail in the discussion forum, she may not describe all aspects she takes into consideration 8 URL: urlhttp://www.librarything.com/topic/129939 when selecting books. This may partly be because she wants to explore di erent options along di erent dimensions and therefore leaves some room for di erent interpretations of her need. Another reason might be that some aspects are not related directly to the topic at hand but may be latent factors that she takes into account with selecting books in general.
What information is available in the LT pro le of the topic starter? First, there are the connections to other members, which come in two kinds, friendship connections and interesting library connections. The former is a connection with another LT member she knows, the latter a connection to an LT member whose personal catalogue she nds interesting. Then there are the group membership connections to the discussion groups on the LT forums. These signal some of her (book-related) interests. Finally, there is the personal catalogue, which contains basic information on the books that the user has added to her personal LT catalogue, when she added each book, as well as any tags and ratings she assigned to each individual book. Basic statistics on the number of connections, books, ratings and tags per user pro le is given in Table 3. In the top row we see the number of days between the topic starter registering to LT and posting the topic on the forum. On average, users build up their pro les for over a year before they ask for book suggestions, although the distribution is skewed. Most book recommendation requests are posted within 215 days of registering to LT, with 54 topics (14%) posted on the day of registration. The latter topics may actually be the reason that the topic starter registered to LT. All frequency distributions, with the exception of group connections, are heavily skewed, with a small number of high end outliers causing the mean to be much higher than the median. The number of groups shown on the pro le is cut-o by LT at 10 groups, which means our pro les miss some group connections for members with more than 10 groups. Most topic starters have a small number of friends and are registered to up to half a dozen discussion groups. Topic starters have around 100 books in their catalogues, almost twice as many tags and only few ratings, although some users have much bigger and more heavily tagged catalogues. There is a lot of variation in the pro les in all aspects mentioned in Table 3. Especially for pro les with many books, tags and connections, recommendation algorithms may need to be tuned to the parts of the pro le that are relevant to the speci c information need of the topic thread.
In the following sections, we describe the procedures for topic selection (Section 3.6) and for suggestion annotation (Section 3.7) procedure, then how we used the annotations to assign relevance values to the suggestions (Section 3.8). 3.6
Three students from the Royal School of Library and Information Science were paid to annotate the narratives of a sample of LT forum topics. We created an interface to help them to 1) select topic threads that are about book search information needs (as opposed to anything else that people discuss on the LT forums), 2) annotate the selected topics describing the type of information need| are users looking for books about a particular topic, in a certain genre, by a certain author, etc.|and 3) annotate the suggestions provided by LT members in the thread. The latter included questions on whether the suggestors have read the suggested books and what their attitudes were towards the books, i.e. positive recommendations vs negative mentions.
An initial set of 9,401 topic threads containing touchstones was made available to the students. Of those, the students annotated over 1100 topics, of which 386 (35%) were identi ed as topics with a book search information need. From the 386, six topics contain no suggestions to any of the books in our A/LT collection. Although all 386 topics were distributed to participants, these 6 topics are discarded from the evaluation.
Topics can have complex information needs, with a combination of multiple relevance aspects. Traditionally, in ir, the focus has been on what a document is about, but in book search there are often many other aspects of relevance. Reuter [7] identi ed 7 general categories of relevance aspects for book search, to which we added the category of known-item information needs: Metadata books with a certain title or by a certain author, editor, illustrator, publisher, in a particular format, or written Accessibility the language, length or level of di culty of a book, Content aspects such as topic, plot, genre, style or comprehensiveness of a book, Engagement books that t a particular mood or interest, or books that are considered high quality or provide a particular reading experience, Novelty books with novel content for the reader, books that are unusual, Familiarity similar to known book or related to previous experience, Socio-Cultural books related the user's socio-cultural background or values, books that are popular or obscure.
Known-item description of known book to identify title and/or author, or published in certain year or period, In the second step, annotators had to indicate which aspects of relevance the topics relate to. Annotators could select multiple relevance categories. For example, for topic 99309 on the politics of muliticulturalism, the topic starter asks for suggestions about a particular topic|i.e., content relevance|but also asks for books that add something new to what he has already read on the topic|i.e., novelty. The distribution of the relevance aspects in the topic set is shown in Table 4. Book search information needs on the LT forums almost always (83% of the 380 topics) contain content aspect. This reinforces the traditional choice in designing best match retrieval models around aspects of document content. Metadata aspects, such as book title and author, are present in 36% of the data set. Other important aspects are familiarity (44%) and engagement (28%) and known-item (24%). Looking for books similar to certain books a user has read is the task of item-based recommender systems, such as that o ered by Amazon ('customers who bought this book also bought...'). It reinforces our interpretation of LT forum book search needs as a task that combines aspects of retrieval and recommendation. Engagement is something that is hard to express in a search engine query. For instance, how can a user search for text books that are funny or high-brow literature that is scary, or books that challenge the reader's own views on a topic? So it is not surprising that readers instead turn to other readers to ask for suggestions. The same holds for read or `heard about' books for which the user only recalls some aspect of the plot, or the some attributes of certain characters. Book search services are of limited use for such known-item topics, but forum members might be able to help out. Accessibility, novelty and sociocultural aspects are less prominent in our sample set. Only 15% of the topics have a single relevance aspect (see Table 5). The vast majority topics represent complex information needs|most topics have 2 or 3 relevance aspects, 144 and 134 topics respectively|with a mean number of 2.47 aspects per topic.
In addition to the above, annotators had to indicate whether the request was for ction, non- ction or both and they had to provide a search query that they would use with a book search engine. We hope that these queries better express the information need than some of the topic thread titles, some of which do not describe the information need at all. Of the 380 topics, 175 (46%) asked for suggestions on ction books, 49 (13%) on non- ction, 55 (14%) on both ction and non- ction, and for 101 topics (27%) the annotator could not tell. The fraction of non- ction topics is lower than last year (49%). We assume that this di erence is caused by giving the annotators the option to indicate they were not sure, whereas in last year's topic selection procedure, we always choose between ction, non- ction or both.
Figure 1 shows an annotated topic (topic 99309) as an example: <topic id="99309"> <query>Politics of Multiculturalism</query> <title>Politics of Multiculturalism Recommendations?</title> <group>Political Philosophy</group> <member>steve.clason</member> <narrative> I'm new, and would appreciate any recommended reading on the politics of multiculturalism. <a href="/author/parekh">Parekh </a>'s <a href="/work/164382">Rethinking Multiculturalism: Cultural Diversity and Political Theory</a> (which I just finished) in the end left me unconvinced, though I did find much of value I thought he depended way too much on being able to talk out the details later. It may be that I found his writing style really irritating so adopted a defiant skepticism, but still... Anyway, I've read <a href="/author/sen">Sen</a>, <a href="/author/rawles">Rawls</a>, <a href="/author/habermas">Habermas</a>, and <a href="/author/nussbaum">Nussbaum</a>, still don't feel like I've wrapped my little brain around the issue very well and would appreciate any suggestions for further anyone might offer.
</narrative> </topic> 3.7
Finally, annotators had to label each book suggestion provided by LT members (including any provided by the topic starter). They had to indicate whether the suggestor has read the book. For the has read question, the possible answers were Yes, No, Can't tell and It seems like this is not a book. They also had to judge the attitude of the suggestor towards the book. Possible answers were Positively, Neutrally, Negatively, Not sure or This book is not mentioned as a relevant suggestion! The latter can be chosen when someone mentions a book for another reason than to suggest it as a relevant book for the topic of request.
The 380 topic threads were started by 378 di erent members in 347 di erent discussion groups, containing altogether 2,568 messages from 1,109 di erent LT min. max. med. mean std.dev members. The distribution of participants, messages and touchstones per topic as shown in Table 6. On average, each thread has 5.39 participants (median is 3), 6.76 messages (median 4) and 15.87 touchstones (median 8). The distributions are all skewed to the right. Most participants contribute touchstones to multiple topics (topics/participant). Messages with touchstones typically contain only one suggestion (median touchstones per message is 1), although some messages contain many touchstones (58 messages contain 10 touchstones or more).
The relationship between having read the book and the suggestor's attitude towards the book is shown in Table 7. In the majority of cases (63%) members suggest books that they have read. It is rather rare for suggestors to state that they have not read a suggested book (8%). More often, suggestors do not reveal whether they have read the book or not (26%). Books mentioned in response to a book search request are often presented in a positive (51%) or neutral (36%) way. Both positive and negative suggestions tend to come from members who have read the books. When books are mentioned in a neutral way, it is often di cult to tell whether the book has been read by the suggestor, although a third of the neutral mentions comes from members who have read the book. There are 162 touchstones that do not refer to books but to author names. In almost all such cases (160 out of 162), annotators skipped the question regarding the attitude of the suggestor towards the book. In the two remaining cases, the annotator indicated the suggestor mentioned the author in a positive way.
How do the attitude labels relate to the books mentioned by the forum members? Are all books mentioned only once or is there a lot of discussion in many threads? Do members agree on which books are good suggestions and which ones are not? Within a thread, books can be mentioned multiple times, by a single member or by multiple members. The 6029 touchstones represent 5092 books. Most books, 4480 or 88% were mentioned only once in the thread, 612 (12%) were mentioned twice or more and 175 (3%) were mentioned three times or more. Of the 4480 books mentioned only once, 2292 (51%) are mentioned positively, 1683 (38%) neutrally, 85 (2%) negatively and 420 (9%) are labelled as either not sure, non-suggestion or were skipped because they were author names misidenti ed as book titles. Books mentioned by forum members tend to be positive suggestions. Of the 612 books mentioned multiple times, attitudes are mostly all positive (230 books, or 38%)) or a mix of positive and neutral (194 or 32%). For 116 books (19%), all attitudes are neutral. Only for 31 books (5%) there is real disagreement|some are positive, some are negative|and for 26 books (4%) attitudes are all negative or mixed neutral and negative. For 15 books (2%) the annotators cannot tell the attitude or indicated the touchstone is not actually a book. In other words, when books are mentioned multiple times, forum members rarely disagree with each other and are mostly positive.
All in all, in response to a book search request, members suggest mostly books they have read and often in a positive way. This supports our choice of using forum suggestions as relevance judgements. 3.8
The annotated suggestions labels were used to determine the relevance value of each book suggestion in the thread. Because some of the books mentioned in the forums are not part of the 2.8 million books in our collection, we rst removed from the suggestions any books that are not in the INEX A/LT collection, which leaves 4572 out of 5092 books (90%). For these 4572 books we derive relevance values.
Forum members can mention books for many di erent reasons. We want the relevance values to distinguish between books that were mentioned as positive recommendations, negative recommendations (books to avoid), neutral suggestions (mentioned as possibly relevant but not necessarily recommended) and books mentioned for some other reason (not relevant at all). We also want to di erentiate between recommendations from members who have read the book they recommend and members who have not. We assume the recommendation to be of more value to the searcher if it comes from someone who has actually read the book. For the mapping to relevance values, we refer to the rst mention of work as the suggestion and subsequent mentions of the same work as replies. We use has read when the forum members have read the book they mention and not read when they have not. Furthermore, we use a number of simplifying assumptions: { When the annotator was not sure if the person mentioning a book has read it, we treat it as not read. We argue that for the topic starter it is not clear there is a di erence in value of such recommendations. { When the annotator was not sure if the person mentioning a book is positive, negative or neutral, we treat it as neutral. Again, for the topic starter there is no clear signal that there is di erence in value. { has read recommendations overrule not read recommendations. Someone who has read the book is in a better position to judge a book than of someone who has not. { positive and negative recommendations neutralise each other. I.e. a positive and a negative recommendation together are the same as two neutral recommendations. { If the topic starter has read a book she mentions, the relevance value is rv = 0. We assume such books have no value as suggestions. { The attitude of the topic starter towards a book overrules those of others.
The system should retrieve books for the topic starter, not for others. { When forum members mention a single work multiple times, we use the last mention as judgement.
With the following decision tree we determine from which forum members want to use the judgements to derive relevance values: 1. Book mentioned by single member ! use that member's judgement 2. Book mentioned by multiple members 2.1 topic starter mentions book 2.1.1 topic starter only suggests neutrally ! use replies of others (2.2) 2.1.1 topic starter suggests positively/negatively ! use starter judgement 2.1.1 topic starter replies ! use starter judgement 2.2 topic starter does not mention book 2.2.2 members who have read the book suggest/reply ! use has read judgements 2.2.2 no member who suggests/replies about a book has read it ! use all judgements Once the judgements per suggested book are determined, we map the annotated judgements to relevance values. The base relevance value of a book that is mentioned in the thread is rv = 2. The values are modi ed according to the following scheme: 1. single judgement 1.1 starter has read judgement ! rv = 0 1.2 starter has not read judgement 1.2.2 starter positive ! rv = 8 1.2.2 starter neutral ! rv = 2 1.2.2 starter negative ! rv = 0 1.3 other member has read judgement 1.3.3 has read positive ! rv = 4 1.3.3 has read neutral ! rv = 2 1.3.3 has read negative ! rv = 0 1.4 other member has not read judgement 1.4.4 not read positive ! rv = 3 1.4.4 not read neutral ! rv = 2 1.4.4 not read negative ! rv = 0 2. multiple judgements 2.1 multiple has read judgements 2.1.1 some positive, no negative ! rv = 6 2.1.1 #positive > #negative ! rv = 4 2.1.1 #positive == #negative ! rv = 2 2.1.1 all neutral ! rv=2 2.1.1 #positive < #negative ! rv = 1 2.1.1 no positive, some negative ! rv = 0 2.2 multiple not read judgements 2.2.2 some positive, no negative ! rv = 4 2.2.2 #positive > #negative ! rv = 3 2.2.2 #positive == #negative ! rv = 2 2.2.2 all neutral ! rv=2 2.2.2 #positive < #negative ! rv = 1 2.2.2 no positive, some negative ! rv = 0 This results in graded relevance values with seven possible values (0, 1, 2, 3, 4, 6, 8). For the 380 topics, there are 4572 relevance values (see Table 8), with 438 from judgements by the topic starter, and 4134 from the judgements of other members. Of these, 3892 are based on single judgements and 242 on judgements from multiple other members. The topic starters contribute only 10% of the relevance values and half of them on books they have not read. The vast majority of values come from single other forum members (3892 or 85%). Also, 3088 relevance values are based on judgements from members who have read the book (67%). 3.9
This year eight teams together submitted 34 runs (see Table 1). The o cial evaluation measure for this task is ndcg@10. It takes graded relevance values into account and concentrates on the top retrieved results. The results are 1484 total mrr
map shown in Table 9. None of the groups used user pro le information for the runs they submitted. The best performing run is run3.all-plus-query.all-doc- elds by RSLIS, which used all topic elds combined against an index containing all available document elds. The second best group is UAms (ILLC) with run inex13SBS.ti qu.bayes avg.LT rating, which uses only the topic titles and moderated query ran against an index containing the title information elds (title, author, edition, publisher, year), user-generated content elds (tags, reviews and awards) and the subject headings and Dewey decimal classi cation titles from the British Library and Library of Congress. The retrieval score of each book was then multiplied by a prior probability based on the Bayesian average of LT ratings for that book. The third group is ISMD, with manual run run ss bsqstw stop words free member free 2013 (to make the table t on the page, it is shortened to run ss bsqstw stop words free member...). This run is generated after removing Book Search Query Stop Words (bsqstw), standard stop words and the member eld from the topics and running against an index where stop words are removed and the remaining terms are stemmed with the Krovetz stemmer. If we ignore the manual runs, ismd is still the third group with the fully automatic run ism run ss free text 2013, which is generated using free text queries on Krovetz stemmed and stop words removed index.
Many teams used similar approaches, with query representations based on a combination of topic elds and indexes based on both professional and usergenerated metadata. It seems that advanced models were implemented that combining topic statements with pro le information or that treat professional metadata di erently from user-generated content. This may be due to the late release the topics and pro les and the submission deadline being early because of changes in the schedule of clef. It is also possible that for most participants this task is felt to be a retrieval task modelled after standard TREC tasks, so there is little attention for recommendation aspects. 4
The goal of this task was to investigate the application of focused retrieval approaches to a collection of digitised books. The scenario underlying this task is that of a user searching for speci c information in a library of books that can provide evidence to con rm or reject a given factual statement. Users are assumed to view the ranked list of book parts, moving from the top of the list down, examining each result. No browsing is considered (only the returned book parts are viewed by users).
Participants could submit up to 10 runs. Each run could contain, for each of the 83 topics (see Section 4.2), a maximum of 1,000 book pages estimated relevant to the given aspect, ordered by decreasing value of relevance.
A total of 11 runs were submitted by 2 groups (6 runs by OUAC (ID=100) and 5 runs by University of Amsterdam (ID=4)), see Table 1. The track builds on a collection of 50,239 out-of-copyright books9, digitised by Microsoft. The corpus is made up of books of di erent genre, including history books, biographies, literary studies, religious texts and teachings, reference works, encyclopaedias, essays, proceedings, novels, and poetry. 50,099 of the books also come with an associated MAchine-Readable Cataloging (MARC) record, which contains publication (author, title, etc.) and classi cation information. Each book in the corpus is identi ed by a 16 character long bookID { the name of the directory that contains the book's OCR le, e.g., A1CD363253B0F403.
The OCR text of the books has been converted from the original DjVu format to an XML format referred to as BookML, developed by Microsoft Development Center Serbia. BookML provides additional structure information, including markup for table of contents entries. The basic XML structure of a typical book in BookML is a sequence of pages containing nested structures of regions, sections, lines, and words, most of them with associated coordinate information, de ning the position of a bounding rectangle ([coords]): <document> <page pageNumber="1" label="PT CHAPTER" [coords] key="0" id="0"> <region regionType="Text" [coords] key="0" id="0"> <section label="SEC BODY" key="408" id="0"> <line [coords] key="0" id="0"> <word [coords] key="0" id="0" val="Moby"/> <word [coords] key="1" id="1" val="Dick"/> </line> <line [...]><word [...] val="Melville"/>[...]</line>[...] </section> [...] </region> [...] </page> [...] </document>
BookML provides a set of labels (as attributes) indicating structure information in the full text of a book and additional marker elements for more complex structures, such as a table of contents. For example, the rst label attribute in the XML extract above signals the start of a new chapter on page 1 (label=\PT CHAPTER"). Other semantic units include headers (SEC HEADER), footers (SEC FOOTER), back-of-book index (SEC INDEX), table of contents (SEC TOC). Marker elements provide detailed markup, e.g., for table of contents, indicating entry titles (TOC TITLE), and page numbers (TOC CH PN), etc.
The full corpus, totaling around 400GB, was made available on USB HDDs. In addition, a reduced version (50GB, or 13GB compressed) was made available for download. The reduced version was generated by removing the word tags and propagating the values of the val attributes as text content into the parent (i.e., line) elements. 9 Also available from the Internet Archive (although in a di erent XML format) In recent years we have had a topic-base of 83 topics, and for 30 of them we have collected relevance judgments using crowdsourcing through Amazon Mechanical Turk [4, 6]. This year we added a new relevance criterium, namely appropriateness. It is not enough that a page con rms or refutes a fact, it should also come from a book that is trusted and of an appropriate genre. For a fact about Darwin's life, a famous biography on Darwin would be a more appropriate source than an obscure textbook on biology. Book pages are judged on two levels: 1) the extent to which a page con rms or refutes the factual claim, determined by how many of the atomic aspects of the claim are con rmed/refuted, and 2) the appropriateness of the book of the page is a part.
Aspect relevance Last year we introduced aspect relevance [6], where complex statements were broken down into atomic parts, which could be judge more easily individually. The overall relevance score of page for the whole statement would be the sum of relevance scores for the atomic parts.
To divide each topic into its primitive aspects (a process we refer to as "aspectisation") we developed a simple web-application with a database back-end, to allow anyone to aspectise topics. This resulted in 30 aspectised topics. The judgements were collected last year. For each page being assessed for con rmation / refutation of a topic, the assessor is presented with a user interface similar to Figure 2
This means that we go from a discrete (con rms / refute / none ) assessment to a graded assessment, where a page may e.g. be assessed by a certain as 33 percent con rming a topic, if one of three aspects is judged as con rmed by him/her for that page. For the current assessment we have prepared 30 topics, for which the number of aspects range from 1 (very simple statements) to 6 per topic with an average of 2,83 aspects per topic.
Appropriateness For this year, the new assessment phase should establish whether the source books for the top-10 pages per topic are appropriate. At the time of writing, the assessment interface is not yet ready, but we expect to run this phase in the summer of 2013. The interface will present assessors with one of the top-10 pages, and provide them with the ability to browse through the rest of the book both via forward/backward buttons, an interactive table of contents in a sidebar as well as a page number box in which they can indicate to which speci c page they want to jump. For each source book of the top-10 pooled pages, the assessor has to judge the appropriateness of the book. This appropriateness score is then propagated to all its pages that are part of the judgement pool. As for the 2012 experiments, one hundred pages were pulled from the top-10 results of participant submissions for each topic. Assessments for each page and statement were collected from three assessors. New for this year, we also ask assessors to assess weather the book a presented page belongs to is appropriate for the task of con rming or refuting the statement. appropriateness is based on the book's genre or topic.
Result publication is awaiting the conclusion of the relevance assessment process. 5
As in previous years, the goal of the SE task was to test and compare automatic techniques for extracting structure information from digitised books and building a hyperlinked table of contents (ToC). The task was motivated by the limitations of current digitisation and OCR technologies that produce the full text of digitised books with only minimal structure markup: pages and paragraphs are usually identi ed, but more sophisticated structures, such as chapters, sections, etc., are typically not recognised.
In 2013, the task was run for the third time as a competition of the International Conference on Document Analysis and Recognition (ICDAR). Full details are presented in the corresponding speci c competition description [3]. This year, the main novelty was that ground truthing was performed by an independent provider. This granted higher consistency and set participants free from ground truthing duties, which were a known drawback of participating to the task. The ground truth for the 2013 competition is already available online10. 10 https://doucet.users.greyc.fr/StructureExtraction/training/ Organization Elsevier EPITA (France) INRIA (France) Microsoft Development Center (Serbia) Nankai University (PRC) NII Tokyo (Japan) University of Caen (France) University of Innsbruck (Austria) University of Wurzburg (Germany) Following the call for participation issued in January 2013, 9 organizations registered. As in previous competitions, several participants expressed interest but renounced due to time constraints. Of the 9 organizations that signed up, 6 submitted runs. This promising increase in active participants (6 out of 9), compared with previous years (4 out of 11), is likely a result of available training data and the removed obligation on creating ground truth ToCs.
As in previous years [
A summary of the performance of all the submitted runs is given in Table 11.
The Structure Extraction task was launched in 2008 to compare automatic techniques for extracting structure information from digitised books. While the construction of hyperlinked ToCs was originally thought to be a rst step on the way to the structuring of digitised books, it turns out to be a much tougher nut to crack than initially expected.
Future work aims to investigate into the usability of the extracted ToCs. In particular we wish to use qualitative measures in addition to the current precision/recall evaluation. The vast e ort that this requires suggests that this can hardly be done without crowdsourcing. We shall naturally do this by building on the experience of the Book Search tasks described earlier in this paper. 6
This was the third year of the Social Book Search Track. This year, the track ran three tasks: the Social Book Search task, the Prove It task and the Structure Extraction task.
RunID MDCS
Nankai Innsbruck Wurzburg
Epita GREYC-run-d GREYC-run-c GREYC-run-a GREYC-run-e GREYC-run-b
Participant
MDCS
Nankai U.
Innsbruck U.
Wurzburg U.
Epita University of Caen University of Caen University of Caen University of Caen University of Caen
The Social Book Search (SBS) task continued which its focus on both the relative value of professional and user-generated metadata and the retrieval and recommendation aspects of the LT forum users and their information needs. The number of active participants has doubled from 4 to 8, suggesting a promising future for the task.
Expanding on the evaluation of the previous two years, we delved deeper into the nature of book search information needs and book suggestions from the LT forums. We annotated both 1) the information needs described by the starters of topic thread there were asking for book suggestions, and 2) the books suggested by LT members in the thread.
We found that most social book search topics have requirements related to the content of the book, such as topic and genre, but that metadata, familiarity and engagement|asking for books by a certain author, books that are similar to a particular (set of) book(s) and books that t a certain mood, interest or quality respectively|are also important aspects. This strengthens and extends our ndings from last year that social book search topics express complex needs that are hard to satisfy with current book search services, but also to speci c for typical recommendation systems.
Another nding in the SBS task is that forum members mostly suggest books they have read although there are also many cases where it is hard to judge from what they write about the books they suggest. When it is clear they have read the books they suggest read, they are mostly positive, supporting our choice of using forum suggestions as relevance judgements. When they suggest books they have not read, or when it is hard to tell, their are often neutral. This could be a signal that suggestions of unread books are closer to traditional topical relevance judgements and suggestions of read books are topic speci c recommendations that satisfy all or most of the complex combination of relevance aspects.
The evaluation has shown that the most e ective systems incorporate the full topic statement, which includes the title of the topic thread, a query provided by the annotator, the name of the forum discussion group, and the full rst message that elaborates on the request. However, the best system is a plain fulltext retrieval system that ignores all user pro le information. It could be that the suggestions by members other than the topic starter favour non-personalised retrieval models and thereby muddle the personalised signal of suggestions supported by the topic starter.
Next year, we plan to shift the focus of the SBS task to the interactive nature of the topic thread and the suggestions and responses given by the topic starter and other members. We are also thinking of a pilot task in which the system not only has to retrieve relevant and recommendable books, but also to select which part of the book description|e.g. a certain set of reviews or tags|is most useful to show to the user, given her information need.
This year the Prove It task changed somewhat by adding the criterium that returned pages should come from reliable sources. That is, authoritative books of the appropriate genre. The assessment phase is yet to start so there are no evaluation results yet. Due to the low number of participants over the last years, the task will probably not run again next year.
The SE task relies on a subset of the 50,000 digitised books of the PI task. In 2013, the participants were to extract the tables of contents of 1,000 books extracted from the whole PI book collection. In previous years, the ground truth was constructed collaboratively by participating institution. For the rst, time in 2013 the ground truth production was performed by an external provider. This centralised construction granted better consistency. In addition, it also validated the collaborative process used since 2009, as the results this year were in line with those of the previous rounds.
The structure extraction task has reached a record high number of active participants, and has for the rst time witnessed an improvement of the state of the art. In future years, we aim to investigate the usability of the extracted ToCs, both for readers in navigating books and systems that index and search parts of books. To be able to build even larger evaluation sets, we hope to experiment with crowdsourcing methods. This may o er a natural solution to the evaluation challenge posed by the massive data sets handled in digitised libraries. Acknowledgments We are very grateful to Toine Bogers for helping us with the topic annotation tool and for recruiting LIS students to be annotators for this year's topic selection and relevance assessments. Recognition (IJDAR), Special Issue on Performance Evaluation of Document Analysis and Recognition Algorithms., 14(1):45{52, 2011. [3] A. Doucet, G. Kazai, S. Colutto, and G. Muhlberger. Overview of the ICDAR 2013 Competition on Book Structure Extraction. In Proceedings of the Twelfth International Conference on Document Analysis and Recognition (ICDAR'2013), page 6, Washington DC, USA, August 2013. [4] G. Kazai, J. Kamps, M. Koolen, and N. Milic-Frayling. Crowdsourcing for book search evaluation: Impact of hit design on comparative system ranking. In Proceedings of the 34th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 205{214. ACM Press, New York NY, 2011. [5] M. Koolen, J. Kamps, and G. Kazai. Social Book Search: The Impact of Professional and User-Generated Content on Book Suggestions. In Proceedings of the International Conference on Information and Knowledge Management (CIKM 2012). ACM, 2012. [6] M. Koolen, G. Kazai, J. Kamps, M. Preminger, A. Doucet, and M. Landoni.
Overview of the INEX 2012 social book search track. In S. Geva, J. Kamps, and R. Schenkel, editors, Focused Access to Content, Structure and Context: 11th International Workshop of the Initiative for the Evaluation of XML Retrieval (INEX'12), LNCS. Springer, 2013. [7] K. Reuter. Assessing aesthetic relevance: Children's book selection in a digital library. JASIST, 58(12):1745{1763, 2007.