In this paper we describe our participation in the INEX 2012 Book Track. The collection enters its second year of age and is composed of Amazon and LibraryThing entries for real books, and their associated user reviews, ratings and tags. Like in 2011, we tried a simple yet e ective approach of reranking books using a social component that takes into account both popularity and ratings. We did experiments using tags as well.
Previous editions of the INEX Book Track focused on the retrieval of real
outof-copyright books [
In 2011, for the books search task, the document collection changed and is now composed of the Amazon pages of real books. IR systems must now search through editorial data and user reviews and ratings for each book, instead of searching through the whole content of the book. The topics were extracted from the LibraryThing1 forums and represent real requests from real users.
Like we already did last year, we used 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 each book, considering the amount of reviews and the ratings. This score is then used to modify the initial ranking obtained by a Markov Random Field baseline that proved to be highly e ective last year. We also used tags to build a pro le for both a query and the books of the collection which we compared to rank the books.
The rest of the paper is organized as follows. The following Section gives an insight into the document collection whereas Section 2 describes the our retrieval framework. Finally, we describe our runs in Section 3.
We used a language modeling approach to retrieval [
Finally, documents are ranked according to the following scoring function: scoreSDM (Q; D) = T
X fT (q; D) q2Q jQj 1 + O X fO(qi; qi+1; D) i=1 jQj 1 + U X fU (qi; qi+1; D) i=1 where the features 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
The basic idea behind this likeliness is that if a book has a lot of reviews and if its ratings are generally good, then it must be a very good book.
L(D) = log(#reviews(D))
Pr2RD r #reviews(D) where RD is the set of all ratings given by the users for the book D, and #reviews(D) is the number of reviews.
We further rerank the books by weighting the previously computed SDM with the likeliness score. The scoring function of a book D given a query Q is thus de ned as follows:
s(Q; D) = L(D) scoreSDM (Q; D) 2.3
We want to represent each query Q by a thematic pro le and rank books according to their relatedness to it. For this rst attempt at using thematic (or topic) relatedness we choosed to rely exclusively on user tags associated with the books in the collection. We consider as a thematic pro le a set of tags weighted according to their signi cance for Q and we call it a tag pro le (TP). As a preprocessing step, a tag pro le is associated to each book in the collection. Tags are weighted according to a classic tf.idf measure (where the tf is the number of users who associated the tag to the book).
The main issue is to estimate a tag pro le for a query. To construct it, inspired by the pseudo relevance feedback method, we summed the pro le of the x top ranked books retrieved by mean of a information retrieval model (more details in runs section). Once the query's tag pro le is build, we can compare book's tag pro le to it with a vector similarity measure like the cosinus.
Finally, books of the collection are ranked according to the similarity of their pro le to the query's one. 3
We submitted 4 runs for the Social Search for Best Books task. We used Indri2 for indexing and searching. We did not remove any stopword and used the standard Krovetz stemmer. mrf-booklike This run is the implementation of the SDM model described in Section 2.1 with the likeliness score.
IOT30 and IT30 Those two runs are based on the tag pro le approach presented in Section 2.3. In this approach four parameters have to be xed : The number x of top ranked books used to build the query's tag pro le, the weight given to each tag in query's pro le, the information retrieval model used to retrieved books and the similarity measure to compare pro les. For both runs, x is xed to 30, Indri's language modeling approach is used and the similarity measure is the cosinus angle between vectors.
The last parameter is the weight given to each tag of the query pro le. For the IOT30 run, the ti tag's weight is compute as the sum of its tf.idf weight in each of the top x books returned by Indri: w(ti) =
X tf:idf (ti; b) b2T opx where b is one the T opx books retrieved.
However, we had the intution that all selected books can not contribute equally to the weight of a tag. So, for the IT30 run, we combine the tf.idf of a tag in a book with the relevance of this book according to the retrieval model used in order to penalize contribution of less relevant books: w(ti) =
X tf:idf (ti; b) score(b; Q) b2T opx
where score(b; Q) is the measure of relevance of the book b according to Indri.
deduce B IT30 30 For the last run we wanted to take advantage of both particularities of mrf-booklike run and a tag pro le based one. We combine the mrf-booklike run to the IT30 run by mean of a logistic regression. We trained a model with two classes (relevant or not) and the book scores predicted by both runs as features. Training instances were the 30 top ranked books returned by each run along with their relevance judgment deduce from 2011 qrels. 4
Recip rank Recall@10 0.2282 0.2071 0.2105 0.2081 0.1890 0.3069 0.3410 0.3584 0.2933 0.2999 0.5398 0.4811 0.4626 0.4524 0.4337 0.1527 0.1659 0.1514 0.1503 0.1426
In this paper we presented our contributions for the INEX 2012 Book Track. We proposed a simple method for reranking books based on their likeliness and an e ective way to take into account user tags. Finally a combination of both methods with a logistic regression approach gives the best results. Results does not allow us to answer on the usefulness of social information for book search despite quite good results.