Jean-Noël Vittaut

Structured multimedia document classification

We propose a new statistical model for the classification of structured documents and consider its use for multimedia document classification. Its main originality is its ability to simultaneously take into account the structural and the content information present in a structured document, and also to cope with different types of content (text, image, etc). We present experiments on the classification of multilingual pornographic HTML pages using text and image data. The system accurately classifies porn sites from 8 European languages. This corpus has been developed by EADS company in the context of a large Web site filtering application.

An algebra for structured queries in bayesian networks

We present a system based on a Bayesian Network formalism for structured documents retrieval. The parameters of this model are learned from the document collection (documents, queries and assessments). The focus of the paper is on an algebra which has been designed for the interpretation of structured information queries and can be used within our Bayesian Network framework. With this algebra, the representation of the information demand is independent from the structured query language. It allows us to answer both vague and strict structured queries.

Learning classification with both labeled and unlabeled data

A key difficulty for applying machine learning classification algorithms for many applications is that they require a lot of hand-labeled examples. Labeling large amount of data is a costly process which in many cases is prohibitive. In this paper we show how the use of a small number of labeled data together with a large number of unlabeled data can create high-accuracy classifiers. Our approach does not rely on any parametric assumptions about the data as it is usually the case with generative methods widely used in semi-supervised learning. We propose new discriminant algorithms handling both labeled and unlabeled data for training classification models and we analyze their performances on different information access problems ranging from text span classification for text summarization to e-mail spam detection and text classification.

Machine learning ranking for structured information retrieval

We consider the Structured Information Retrieval task which consists in ranking nested textual units according to their relevance for a given query, in a collection of structured documents. We propose to improve the performance of a baseline Information Retrieval system by using a learning ranking algorithm which operates on scores computed from document elements and from their local structural context. This model is trained to optimize a Ranking Loss criterion using a training set of annotated examples composed of queries and relevance judgments on a subset of the document elements. The model can produce a ranked list of documents elements which fulfills a given information need expressed in the query. We analyze the performance of our algorithm on the INEX collection and compare it to a baseline model which is an adaptation of Okapi to Structured Information Retrieval.

Segmented Images Colorization Using Harmony

In the colorization process, artists fill images regions with colors following harmony rules. Some region colorizations are more significant than others. In this paper we introduce a model for computer assisted colorization of segmented images using Ittens proportion contrast. Our model is able to process images composed of totally blank or partially filled regions. Though, each region must be filled with only one color. The user defines sectors representing colors on the chromatic hue wheel and may redefine each colors proportion. Then a set of colors is selected to compute the desired harmony. To proportionally colorize an image, already existing colors can be taken into account or the harmony can be applied only to blank regions. Due to the fact that regions cannot be divided, exact proportion contrast colorization of segmented images is a difficult algorithmic problem. In order to solve this problem we use different strategies and propose different methods to fill the regions. If they exist, exact solutions are proposed, otherwise different approximation methods are used to determinate combinations of colors-regions close to the defined proportions. Once the image is processed, the user can adapt hue (in the chosen sector), saturation and value of each region to enhance the result. Our model is flexible and specifically designed to help artists, illustrators, comics creators or any other user to automatically colorize (partially or totally) their images.

Machine learning ranking and INEX’05

We present a Machine Learning based ranking model which can automatically learn its parameters using a training set of annotated examples composed of queries and relevance judgments on a subset of the document elements. Our model improves the performance of a baseline Information Retrieval system by optimizing a ranking loss criterion and combining scores computed from doxels and from their local structural context. We analyze the performance of our algorithm on CO-Focussed and CO-Thourough tasks and compare it to the baseline model which is an adaptation of Okapi to Structured Information Retrieval.

Fast instantiation of GGP game descriptions using prolog with tabling

We present a method to instantiate game descriptions used in General Game Playing with a Prolog interpreter using tabling. Instantiation is a crucial step for speeding up the interpretation of game descriptions and increasing the playing strength of general players. Our method allows us to ground almost all of the game descriptions present on the GGP servers in a time that is compatible with the common time settings of the GGP competition. It instantiates more rapidly than previous published methods.

Supervised and Semi-supervised Machine Learning Ranking

We present a Semi-supervised Machine Learning based ranking model which can automatically learn its parameters using a training set of a few labeled and unlabeled examples composed of queries and relevance judgments on a subset of the document elements. Our model improves the performance of a baseline Information Retrieval system by optimizing a ranking loss criterion and combining scores computed from doxels and from their local structural context. We analyze the performance of our supervised and semi-supervised algorithms on CO-Focussed and CO-Thourough tasks using a baseline model which is an adaptation of Okapi to Structured Information Retrieval.

A General Approach of Game Description Decomposition for General Game Playing

We present a general approach for the decomposition of games described in the Game Description Language (GDL). In the field of General Game Playing, the exploration of games described in GDL can be significantly sped up by the decomposition of the problem in sub-problems analyzed separately. Our program can decompose game descriptions with any number of players while addressing the problem of joint moves. This approach is used to identify perfectly separable sub-games but can also decompose serial games composed of two subgames and games with compound moves while avoiding, unlike previous works, to rely on syntactic elements that can be eliminated by simply rewriting the GDL rules. We tested our program on 40 games, compound or not, and we can decompose 32 of them successfully in less than 5 s.

Efficient grounding of game descriptions with tabling

We present a method to instantiate game descriptions used in General Game Playing with the tabling engine of a Prolog interpreter. Instantiation is a crucial step for speeding up the interpretation of the game descriptions and increasing the playing strength of general game players. Our method allows us to ground almost all of the game descriptions present on the GGP servers in a time that is compatible with the common time settings of the GGP competition. It instantiates descriptions more rapidly than previous published methods.