In this paper, we present our contribution in Suggestion Track at the Social Book Search Lab. This track aims to develop test collections for evaluating ranking e ectiveness of book retrieval and recommender systems. In our experiments, we combine the results of Sequential Dependence Model (SDM) and the books information that includes the price, the number Of P ages and the publication Date. We also expand topics' queries by the similar books information to improve the recommendation performance.
The Social Book Search (SBS) Tracks [
Social Book Search Lab includes the following tracks: Suggestion Track, Interactive Track and Mining Track. Our work is on Suggestion Track, which suggests a list of the most relevant books according to the request provided by the user. Since 2011, for the social books search task, the document provided is a collection of 2.8 million records containing professional metadata (Amazon1) extended with user-generated content and social metadata (LibraryThing2). In addition, a set of 113,490 anonymous users pro les is provided from LibraryThing (LT). Therefore, Information Retrieval (IR) Systems must search through editorial data, user reviews and ratings for each book, instead of searching through the whole content of the book. The topics provided each year are extracted from the LibraryThing forums and by represent real requests from real users.
Our participation in 2011 and 2012 was based on re-ranking books using
social component such as popularity and ratings [
This year's participation is through an IR system based on 3 main steps: { We expand the topic queries using the similar books information, since the topics contain books titles mentioned by the user as similar or example books to those he seeks. { We apply a re-ranking method using a score calculated of books information including the price, the number Of P ages and the publication Date. { We apply these methods on Amazon book collection and on the users pro les collection.
For our participation in SBS 2016, we submitted 4 runs in which we applied the previously mentioned steps. The rest of this paper is organized as follows. The following section describes the data processing and indexing. In section 3, we have the description of our retrieval framework. In section 4, we describe the submitted runs. Finally, we present the obtained results in section 5. 2
We use, in addition to the Amazon book Collection, the users pro les Collection provided by SBS Lab track which contains the cataloguing transactions of 113,490 users. The cataloguing transactions of a user is a list of information concerning the books read by the user. Each transaction is represented by a row, where each row contains eight columns; user, book, author, book title, publication year, month in which the user added that book, rating and a set of tags assigned by this user to this book. From the users pro les, we create for each book an XML le with all its information. An example is illustrated in the following XML code of Figure 1.
For indexing the Amazon book collection, we take all the tags of the XML les identi ed by the ISBNs. And for indexing users pro les collection, we take all the tags of the created XML les identi ed by the LibraryThingID. Also, we use the following Indri4 indexing parameters: P orter Stemmer and Stop W ords Removal.
3
Query Expansion by example books information To build our queries we use mainly the title of the query and the information of similar example books mentioned by the user in the topic. Also, we use the tags of these similar books extracted from the users pro les collection for query expansion. The XML code in Figure 2 illustrates an example of adding similar book tags for query expansion. 3.2
Sequential Dependence Model
SDM relies on the idea of integrating multi word phrases by considering a
combination of query terms with proximity constraints such as: single term features
(standard unigram language model features, fT ), exact phrase features (words
appearing in sequence, fO) and unordered window features (require words to be
close together, but not necessarily in an exact sequence order, fU ) [
jCj cf#uw8(qi;qi+1) jCj
in document D.
'qi qi+1' in document D.
window 'qi qi+1' (span=8) in document D.
And the documents are ranked according to the below scoring equation, Equation 1:
SDM (Q; D) =
T Pq2Q fT (q; D) + O PjiQ=j1 1 fO(qi; qi + 1; D) + U PjiQ=j1 1 fU (qi; qi + 1; D) (1)
We used the Equation 1 with feature weights set to T = 0.85, O = 0.1 and U = 0.05, like previous participation years. We applied this model to the queries using Indri 5.4 4 Query Language 5. An example of Indri Query Language is in Figure 3.
Combination of Retrieval System output and books' information
We combine the results of SDM model with a sum of normalized scores, which
we calculate from the book's price, publication Date and number Of P ages.
And since the combined values are of di erent weighting, we use the maximum
and minimum scores according to Lees formula [
The scores of SDM model and books information have di erent levels of retrieval e ectiveness, thus it is necessary to weigh scores depending on their overall performance. We used an interpolation parameter ( ) that varies in testing for the goal of achieving the best interpolation that provides better retrieval e ectiveness, as shown in the Equation 3.
SDM bookInf o = :(SDM (Q; D)) + (1 ):(bookInf o(D)) (3)
After several testings on 2015 SBS topics 6, is set to 0.55 with the best result. bookInf o(D) is calculated by a normalized score of the values of price only, since the price alone obtains the best result on 2015 SBS topics compared to the values of price, publication Date and number Of P ages combined. In Table 2, an example of our tests showing a modest but still an increase in the results when combining books prices to the equation with = 0.55.
We submit 4 runs for the SBS Suggestion Track: Run1 ExeOrNarrativeNSW Collection: We concatenate the title of the topic and the similar books elds (title, author and tags), then perform a retrieval using the SDM model. But since not all topics have example books, in this case we concatenate the title and the narrative elds of the topic after removing the Stop W ords from the narrative eld. This run is applied on Amazon book collection.
Run2 ExeOrNarrativeNSW UserPro le : This run is same as Run1 but it is applied on users pro les collection.
Run3 ExeOrNarrativeNSW Collection AddData : In this run, we combine the books price normalized score to the results of Run1.
Run4 ExeOrNarrativeNSW UserPro le AddData : Also in this run, we combine the books price normalized score to the results of Run2. 5
In this paper, we present our contribution for the Suggestion Track of Social Book Search Lab. In the 4 submit runs, we use SDM retrieval model and we extend the query by the similar books information (title, author and tags). We apply the retrieval on Amazon book collection, and on users pro les collection. We combine the results of the retrieval system (SDM) with the normalized score of the books prices. The best result is achieved by using SDM retrieval model with the extended query on Amazon book Collection. We should note that the topics of SBS 2015 had a eld named mediated query, which contained the key words of the user's request (from eld narrative). The mediatedquery eld is used in our testing on SBS 2015 topics and helped to increase the results. But since this eld is not in the topics of SBS 2016, we had to use the narrative eld which contains many useless information that e ect negatively the information research. Thus, to increase the results for future participation, we must work on extracting only the key words from the narrative eld to be used in the query, and eliminate any noise information.
Acknowledgments. This work was supported by the French program Investissements dAvenir Equipex "A digital library for open humanities" of OpenEdition.org.