Loïc Maisonnasse

Multiplying Concept Sources for Graph Modeling

The main idea in this paper is to incorporate medical knowledge in the language modeling approach to information retrieval (IR). Our model makes use of the textual part of ImageCLEFmed corpus and of the medical knowledge as found in the Unified Medical Language System (UMLS) knowledge sources. The use of UMLS allows us to create a conceptual representation of each sentence in the corpus. We use these representations to create a graph model for each document. As in the standard language modeling approach, we evaluate the probability that a document graph model generates the query graph. Graphs are created from medical texts and queries, and are built for different languages, with different methods. After developing the graph model, we present our tests, which involve mixing different concepts sources (i.e. languages and methods) for the matching of the query and text graphs. Results show that using language model on concepts provides good results in IR. Multiplying the concept sources further improves the results. Lastly, using relations between concepts (provided by the graphs under consideration) improves results when only few conceptual sources are used to analyze the query.

Model Fusion in Conceptual Language Modeling

We study in this paper the combination of different concept detection methods for conceptual indexing. Conceptual indexing shows effective results when large knowledge bases are available. But concept detection is not always accurate and errors limit interest of concept usage. A solution to solve this problem is to combine different concept detection methods. In this paper, we investigate several ways to combine concept detection methods, both on queries and documents, within the framework of the language modeling approach to IR. Our experiments show that our model fusion improves the standard language model by up to 17% on mean average precision.

Revisiting the dependence language model for information retrieval

In this paper, we revisit the dependence language modelfor information retrieval proposed in [1], and show that thismodel is deficient from a theoretical point of view. We thenpropose a new model, well founded theoretically, for integratingdependencies between terms in the language model.This new model is simpler, yet more general, than the oneproposed in [1], and yields similar results in our experiments,on both syntactic and semantic dependencies.

Visual graph modeling for scene recognition and mobile robot localization

Image retrieval and categorization may need to consider several types of visual features and spatial information between them (e.g., different point of views of an image). This paper presents a novel approach that exploits an extension of the language modeling approach from information retrieval to the problem of graph-based image retrieval and categorization. Such versatile graph model is needed to represent the multiple points of views of images. A language model is defined on such graphs to handle a fast graph matching. We present the experiments achieved with several instances of the proposed model on two collections of images: one composed of 3,849 touristic images and another composed of 3,633 images captured by a mobile robot. Experimental results show that using visual graph model (VGM) improves the accuracies of the results of the standard language model (LM) and outperforms the Support Vector Machine (SVM) method.

LIG at ImageCLEF 2008

This paper describes the work of the LIG for ImageCLEF 2008. For ImageCLEFPhoto, two non diversified runs (text only and text + image), and two diversified runs were officially submitted. We add in this paper results on image only runs. The text retrieval part is based on a language model of Information Retrieval, and the image part uses RGB histograms. Text+image results are obtained by late fusion, by merging text and image results. We tested three strategies for promoting diversity using date/location or visual features. Diversification on image only runs does not perform well. Diversification on image and text+image outperforms non diversified runs. In a second part, this paper describes the runs and results obtained by the LIG at ImageCLEFmed 2008. This contribution incorporates knowledge in the language modeling approach to information retrieval (IR) through the graph modeling approach proposed in [4]. Our model makes use of the textual part of the corpus and of the medical knowledge found in the Unified Medical Language System (UMLS) knowledge sources. And the model is extended to combine different graph detection methods on queries and documents. The results show that detection combination improves the performances.

Spatial relationships in visual graph modeling for image categorization

In this paper, a language model adapted to graph-based representation of image content is proposed and assessed. The full indexing and retrieval processes are evaluated on two different image corpora. We show that using the spatial relationships with graph model has a positive impact on the results of standard Language Model (LM) and outperforms the baseline built upon the current state-of-the-art Support Vector Machine (SVM) classification method.

Analysis combination and pseudo relevance feedback in conceptual language model

This paper presents the LIRIS contribution to the CLEF 2009 medical retrieval task (i.e. ImageCLEFmed). Our model makes use of the textual part of the corpus and of the medical knowledge found in the Unified Medical Language System (UMLS) knowledge sources. As proposed in [6] last year, we used a conceptual representation for each sentence and we proposed a language modeling approach. We test two versions of conceptual unigram language model; one that use the log-probability of the query and a second one that compute the Kullback-Leibler divergence. We used different concept detection methods and we combine these detection methods on queries and documents. This year we mainly test the impact of the use of additional analysis on queries. We also test combinations on French queries where we combine translation and analysis, in order to solve the lack of French terms in UMLS, this provide good results close from the English ones. To complete these combinations we proposed a pseudo relevance method. This approach use the n first retrieve documents to form one pseudo query that is used in the Kullback-Leibler model to complete the original query. The results of this approach show that extending the queries with such an approach improves the results.

Integration of spatial relationships in visual language model for scene retrieval

In this paper, we describe a method to use a graph-based language modeling approach for image retrieval and image categorization. We first mapped image regions to induced concepts and then spatial relationships between these regions to build a graph representation of images. Our method allows to deal with different scenarii, where isolated images or groups of images are used for training and testing. The results obtained on an image categorization problem comprising of 3849 images from 101 landmarks of Singapore show that (a) the procedure to automatically induce concepts from an image is effective, and (b) the use of spatial relationships, in addition to concepts, for representing an image content helps improve the classifier accuracy. This approach is the first one, to our knowledge, to present a complete extension of the language modeling approach from information retrieval to the problem of graph-based image categorization and retrieval.

Modèle de graphe et modèle de langue pour la reconnaissance de scènes visuelles.

Dans cet article, nous decrivons une methode pour utiliser un modele de langue sur des graphes pour la recherche et la categorisation d’images. Nous utilisons des regions d’images (associees automatiquement a des concepts visuels), ainsi que des relations spatiales entre ces regions, lors de la construction de la representation sous forme de graphe des images. Notre methode gere differents scenarios, selon que des images isolees ou groupees sont utilisees comme base d’apprentissage ou de test. Les resultats obtenus sur un probleme de categorisation d’images montrent (a) que la procedure automatique qui associe les concepts a une image est efficace, et (b) que l’utilisation des relations spatiales, en plus des concepts, permet d’ameliorer la qualite de la classification. Cette approche presente donc une extension du modele de langue classique en recherche d’information pour traiter le probleme de recherche et de categorisation d’images non annotees, representees par des graphes.

MRIM-LIG at ImageCLEF 2009: robotvision, image annotation and retrieval tasks

This paper describes mainly the experiments that have been conducted by the MRIM group at the LIG in Grenoble for the the ImageCLEF 2009 campaign, focusing on the work done for the Robotvision task. The proposal for this task is to study the behaviour of a generative approach inspired by the language model of information retrieval. To fit with the specificity of the Robotvision task, we added post-processing in a way to tackle with the fact that images do belong only to several classes (rooms) and that image are not independent from each others (i.e., the robot cannot in one second be in three different rooms). The results obtained still need improvement, but the use of such language model in the case of Robotvision is showed. Some results related to the Image Retrieval task and the Image annotation task are also presented.