UNT at ImageCLEF 2010: CLIR for Wikipedia Images Page: 2
The following text was automatically extracted from the image on this page using optical character recognition software:
(Muller, et al., 2009). There has been a lot on emphasis on trying to improve the
current Content-Based Image Retrieval (CBIR) using automatically extracted visual
features to match sample images given in the official topics. However, our own
research as well as the results from other participants has shown that the most
successful approach to solve the image retrieval problem relies on high quality text
retrieval (Muller, et al., 2009; Ruiz M. E., 2006; Ruiz & Neveol, 2007). The
combination of visual and textual features has proven to contribute to small
improvements in mean average precision (MAP) which is the standard measure that is
used in CLEF to compare system performance.
One of the key issues that needs to be explored is to find approaches that can
contribute more to solve the retrieval problem when the given data collection contains
annotations generated in multiple languages. For the current Wikipedia retrieval task
this is an important issue since the collection has a relatively even distribution of
annotations in three languages (English, French and German). The CLIR problem has
been solved using several methods but the most commonly used approach consists of
translating the text of the user query to the language of the document, performing
monolingual retrieval in each language and then combining the results of several
monolingual runs. However, this approach has two main potential challenges:
1. Use of machine translation on short queries can be difficult due to the loss
of context and finding appropriate disambiguation for automatic
2. Finding optimal parameters to adjust the mechanism to merge results from
multiple monolingual runs is challenging. Moreover, these optimal
parameters can change from one collection to another making it hard to
find a general optimal set of parameters.
We decided to explore a solution that translates all the documents to a single
language and perform the retrieval in that language only. We recognize that this is an
expensive solution that might not work for a general CLIR problem. However, for
image retrieval it is a viable option due to the relatively short length of image captions
(compared to the full text associated with the images in an article in Wikipedia). Also,
image captions usually contain enough contextual information to allow appropriate
translation disambiguation. The translation of captions can be achieved relatively fast
using MT translation systems that are freely available on the Internet such as Google
Translation. This also reduces the CLIR problem to a simple monolingual translation
for which the technology is more stable, and there is no need to deal with merging the
results that come from different collections.
We used the Indri/Lemur Retrieval system to index our collection using standard
Krovetz stemming and the standard Language Model implemented in Lemur ( Lemur
Here’s what’s next.
This paper can be searched. Note: Results may vary based on the legibility of text within the document.
Tools / Downloads
Get a copy of this page or view the extracted text.
Citing and Sharing
Basic information for referencing this web page. We also provide extended guidance on usage rights, references, copying or embedding.
Reference the current page of this Paper.
Ruiz, Miguel E.; Chen, Jiangping; Pasupathy, Karthikeyan; Chin, Pok & Knudson, Ryan. UNT at ImageCLEF 2010: CLIR for Wikipedia Images, paper, September 2010; (https://digital.library.unt.edu/ark:/67531/metadc96836/m1/2/: accessed May 20, 2019), University of North Texas Libraries, Digital Library, https://digital.library.unt.edu; crediting UNT College of Information.