: In this paper, we present our contribution in INEX 2013 Social Book Search Track. This track aim to explore social information (users reviews, ratings, etc...) for the libraryThing and Amazon collections of real books. In our submissions for SBSTrack, we rerank books by combining the Sequential Dependence Model (SDM) and the use of social component that takes into account both ratings and helpful votes.
Previous editions of the INEX Book Track focused on the retrieval of real out- of-copyright books [1]. These books were written almost a century ago and the collection consisted of the OCR content of over 50, 000 books. The topics and the books of the collection have a di erent vocabulary and writing style. Information Retrieval systems had di culties to found relevant information, and assessors had di culties judging the documents.
The document collection is composed of the Amazon 1 pages of real books. IR must search through editorial data, user reviews and ratings for each book, instead of searching through the whole content of the book. The topics were extracted from LibraryThing 2 forums and represent real request from real users.
We have chosen to use a Language Modeling approach to retrieval. For our recommendation runs, we used the reviews and the ratings attributed to books by Amazon users. We computed a \social score" for 1http://www.amazon.com/ 2http://www.librarything.com/ each book, considering the amount of reviews and the ratings. This score was then interpolated with scores obtained by a Marcov Random Field (MRF) baseline. We also used the \helpfulvotes" and \totalvotes" values for each rating given by users to modify the ranking obtained by the combination of social and MRF scores.
The rest of the paper is organized as follows. The following Section gives an insight into the document collection whereas 2 describes the our retrieval framework. Finally, we describe our runs in 3. 2 2.1
Sequential Dependence Model We used a language modeling approach to retrieval [2]. We use Metzler and Croft's Markov Random Field (MRF) model [3] to integrate multi word phrases in the query. Speci cally, we use the Sequential Dependence Model (SDM), which is a special case of MRF. In this model three features are considered: 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 ).
Finally, documents are ranked according to the following scoring function:
SDM (Q; D) =
T X fT (q; D)
q2Q + O + U i=1 i=1 jQj 1 X fO(qi; qi+1; D) jQj 1 X fU (qi; qi+1; D) where the feature weights are set according to the author's recommendation ( T = 0.85, O= 0.1, U = 0.05). fT , fO and fU are the log maximum likelihood estimates of query terms in document D, computed over the target collection with a Dirichlet smoothing. 2.2
Modeling book likeliness We modeled book likeliness basing on the following idea: more the number of reviews it has, more interesting it is reading it (it may not be a good or popular book but a book that has a high impact).
Likeliness(D) =
Pr2RD r jReviewsDj where RD is the set of all ratings given by the users for the book D, and jReviews Djis the number of reviews.
We further rerank books according to a linear interpolation of the
previously computed SDM score with the likeliness score, using a
coefcient ( ) to control the in uence of each model. The scoring function
of a book D given a query Q is thus de ned as follows:
SDM Likeliness(Q; D) =
(SDM (Q; D)) + (1
) (Likeliness(D))
where is a constant set according to previous results
Modeling usefulness of ratings' books Into the collection of books, we have a rating for each review given by users, the rating value can or cannot be useful depending on user votes. we have chosen to weight the value of rating with the value of helpful votes according to this formula:
U sef ulness(D) =
P r2RD;t2TD;h2HD r jReviewsDj ( ht ) where RD , TD , HD are respectively, the sets of all ratings, totalvotes and helpfulvotes given by the users for the book D, and jReviews Djis the number of reviews.
We further rerank books according to a linear interpolation of the previously computed SDM score with the usefulness score, using a coefcient ( ) to control the in uence of each model. The scoring function of a book D given a query Q is thus de ned as follows: SDM U sef ulness(Q; D) = (SDM (Q; D))+(1
) (U sef ulness(D))
where is a constant set according to previous results
Run We submitted 3 runs for the Social Book Search Task. We used Indri 3 for indexing and searching. We did not remove any stopword and used the standard Krovetz stemmer. Only query part of the topic has been used for the three runs.
SDM run : This run is the implementation of the Sequential Dependence Model (SDM) described in Section 2.1.
SDM Rating run : This run combine the implementation of the Sequential Dependence Model and the use of social information which is the \Ratings" given by users. Description is given in Section 2.2
SDM HV run : For the last run we combine the implementation of the Sequential Dependence Model and the use of social information which are \ratings", \helpful votes" and \total votes" given by users. We weighted the value of \rating" with the rate of helpful votes as presented in Section 2.3. 4
This paper presents our contributions on the INEX 2013 Social Book Search Track. We proposed a simple method for reranking books based on their likeliness and an e ective way to take into account user helpful votes. Finally we combine both methods with a linear interpolated function.
We are disappointed with the o cial results this year (nDCG@10 = 0.0571 for the baseline \SDM run") compared to those obtained last year with the the approach that was our baseline this year (nDCG@10 = 0,1295) and we seek for explanations of a software problem. On the other side, the proposed extensions (nDCG@10 = 0.596 for \SDM Rating run" and nDCG@10 = 0.0576 for \SDM HV run") improved the results of the baseline.
This work was supported by the French program \Investissements d'Avenir - Developpement de l'Economie Numerique" under the project InterTextes #O14751-408983.