In this paper we will describe Berkeley's approach to the ImageCLEF Wikipedia Retrieval task for 2010. Our approach to this task was primarily to use text-based searches on the contents of the Wikipedia image metadata records. In addition we submitted one run using a database derived from the provided “bag.xml” set of 5000 descriptor “words” for each image and query example images. We had also intended to combine this one image-based approach to other image-based approaches and to the text-based approaches using fusion methods, but were unable to complete the coding in time. We submitted 8 runs for ImageCLEF Wikipedia Retrieval this year, of which 6 where monolingual English, German and French with differing search areas in the metadata record, one was multilingual and the remaining one was image-based using the data derived from bag.xml file. Our best performing run was ranked 24th among the 127 submitted runs by all participants with a MAP of 0.2014, while the image-only approach was ranked dead last (one wonders, in fact, if random results might have done better).
This paper discusses the retrieval methods and evaluation results for Berkeley’s participation in the ImageCLEF Wikipedia Retrieval task. This year we used primarily text-based retrieval methods for ImageCLEF Wikipedia Retrieval, but also attempted to use some of the supplied image-derived information. We did manage to submit a single image-only run, but were not able to use it in combination with our text-based approach (unfortunately we were not able to complete the merger software in time for official submissions, although we hope to have some combined runs to report later).
This year Berkeley submitted 8 runs, of which 2 where English Monolingual, 2 German Monolingual, and 2 were French monolingual. The remaining runs included one multilingual run (using the English German and French topic text and the entire metadata record as a search target), and a single image-based run derived from the “bag.xml” file provided with the database.
This paper first describes the retrieval methods used, including our blind feedback method for text, followed by a discussion of our official submissions and the methods used for query expansion. Finally we present some discussion of the results and our conclusions. 2
(Note, this section repeats information provided in our 2006 ImageCLEF
Notebook paper, since the basic retrieval algorithms used and the approaches to
indexing the content have not been changed since then. This same algorithm and
approach have been used in a wide variety of cross-language retrieval experiments
in both CLEF and NTCIR (see, for example, [
The basic form and variables of the Logistic Regression (LR) algorithm used
for all of our text-based submissions was originally developed by Cooper, et al.
[
For all of our ImageCLEF submissions this year we used a version of the
Logistic Regression (LR) algorithm that has been used very successfully in
CrossLanguage IR by Berkeley researchers for a number of years[
S i=1
bisi
P (R | Q, D) = 1 + elog O(R|Q,D)
elog O(R|Q,D)
developed by Cooper et al. [
p(R|C, Q) 1 − p(R|C, Q) = log p(R|C, Q) p(R|C, Q) 1
|Qc| qtfi |Qc| + 1 i=1 ql + 35 1 1 |Qc| |Qc| |Qc| + 1 i=1 |Qc| + 1 i=1 log log
tfi cl + 80 ctfi
Nt where C denotes a document component (i.e., an indexed part of a document which may be the entire document) and Q a query, R is a relevance variable, p(R|C, Q) is the probability that document component C is relevant to query
Q, p(R|C, Q) the probability that document component C is not relevant to query
Q, which is 1.0 - p(R|C, Q) |Qc| is the number of matching terms between a document component and a query, qtfi is the within-query frequency of the ith matching term, tfi is the within-document frequency of the ith matching term, ctfi is the occurrence frequency in a collection of the ith matching term, ql is query length (i.e., number of terms in a query like |Q| for non-feedback situations), cl is component length (i.e., number of terms in a component), and Nt is collection length (i.e., number of terms in a test collection). ck are the k coefficients obtained though the regression analysis.
If stopwords are removed from indexing, then ql, cl, and Nt are the query length, document length, and collection length, respectively. If the query terms are re-weighted (in feedback, for example), then qtfi is no longer the original term frequency, but the new weight, and ql is the sum of the new weight values for the query terms. Note that, unlike the document and collection lengths, query length is the “optimized” relative frequency without first taking the log over the matching terms.
The coefficients were determined by fitting the logistic regression model
specified in log O(R|C, Q) to TREC training data using a statistical software package.
The coefficients, ck, used for our official runs are the same as those described
by Chen[
In addition to the direct retrieval of documents using the TREC2 logistic
regression algorithm described above, we have implemented a form of “blind relevance
feedback” as a supplement to the basic algorithm. The algorithm used for blind
feedback was originally developed and described by Chen [
Blind relevance feedback is typically performed in two stages. First, an initial search using the original topic statement is performed, after which a number of terms are selected from some number of the top-ranked documents (which are presumed to be relevant). The selected terms are then weighted and then merged with the initial query to formulate a new query. Finally the reweighted and expanded query is submitted against the same collection to produce a final ranked list of documents. Obviously there are important choices to be made regarding the number of top-ranked documents to consider, and the number of terms to extract from those documents. For ImageCLEF this year, having no prior data to guide us, we chose to use the top 10 terms from 10 top-ranked documents. The terms were chosen by extracting the document vectors for each of the 10 and computing the Robertson and Sparck Jones term relevance weight for each document. This weight is based on a contingency table where the counts of 4 different conditions for combinations of (assumed) relevance and whether or not the term is, or is not in a document. Table 1 shows this contingency table.
The relevance weight is calculated using the assumption that the first 10 documents are relevant and all others are not. For each term in these documents the following weight is calculated: wt = log
Rt R−Rt
Nt−Rt N−Nt−R+Rt (3)
The 10 terms (including those that appeared in the original query) with the highest wt are selected and added to the original query terms. For the terms not in the original query, the new “term frequency” (qtfi in Equation 3 above) is set to 0.5. Terms that were in the original query, but are not in the top 10 terms are left with their original qtfi. For terms in the top 10 and in the original query the new qtfi is set to 1.5 times the original qtfi for the query. The new query is then processed using the same LR algorithm as shown in Equation 3 and the ranked results returned as the response for that topic. 3
Approaches for ImageCLEF Wikipedia Retrieval In this section we describe the specific approaches taken for our official submitted runs for the ImageCLEF Wikipedia Retrieval task. First we describe the indexing and term extraction methods used, and then the search features we used for the submitted runs.
Cheshire II system uses the XML structure of documents and extracts selected portions of those record for indexing and retrieval. In our submitted runs this year we used separate indexes for each of the languages (English, German and French) as well as a global index combining all elements of the metadata records.
Name Description
Content Tags
Used
In earlier work we used the Berkeley Blobworld algorithms [
In looking at the image feature data provided with the collection, we realized that a similar approach might be attempted with that data, so as a last-minute attempt, we tokenized the 5000 element “bag” vectors, using a very simple approach (probably too simple, considering the rather terrible results) and used the basic TREC2 algorithm (without blind feedback) to rank the results. The same approach was used on the provided sample images for the topics to provide the queries for processing. 3.3
Searching the ImageCLEF Wikipedia collection used Cheshire II scripts to parse the topics and submit a query to the system using the topic title in a particular language (or for all of the languages in the multilingual case). Depending on settings in the script, the queries were run against the language specific indexes, or the entire document. The TREC2 algorithm with blind feedback used the top 10 terms from the 10 top-ranked documents in the initial retrieval for the blind feedback.
Our single image-based run used the tokenized features derived from the example images to search the tokenized feature vector indices. The summary results (as Mean Average Precision) for all of our official submitted runs are shown in Table 3, the Recall-Precision curves for the text-based runs are also shown in Figures 1 (for monolingual) and 2 (for multilingual). In Figures 1 and 2 the names are abbrevated as indicated in the “Abbrev.” column. 3. Our officially submitted runs using text retrieval with blind feedback did not perform as well as some of the other participants’ text-only runs and definitely lagged behind the best performing mixed image and text runs. What was also apparent (both from the results and examination of the database itself) was that the multilingual character of the data was very spotty. Most of the metadata entries did not include all three target languages, and many records contained no text descriptions at all other than a caption in a single language, or in some cases terms in the image name itself. Thus, our runs that targetted the entire record did much better than those attempting to access the language-tagged text. Another interesting point is that simply combining the topic titles for each language was our best performing run. This is interesting primarily in comparison with previous CLEF multilingual tasks in other track, where this approach usually made results worse than monolingual approaches. It may be that the sparseness of the metadata records and uneven distribution of language use favors the multilingual approach.
As note above, our single image-only submission was a dismal failure. However, we hope to be able to further test using image element summarization and tokenization, but to do that effectively we first need to know much more about how the supplied feature vectors were created and what each element in those vectors represents. Considering the only real documentation is pointers to journal papers, where the data format is not described at all, it is probably not surprising that our last minute approach made some wrong assumptions and choices in representing the image features. We also still hope to be able to revive the blobworld sofware and use it for future experiments, but at present that is dependent on either getting funding to re-license Matlab, or completing the rewrite of the software in C or some other high-level language.