Résumé: In this paper, we describe our system that participated in SemEval-2013, Task 2 (sentiment analysis in Twitter). Our approach consists of adapting Naive Bayes probabilities in order to take into account prior knowledge (represented in the form of a sentiment lexicon). We propose two different methods to efficiently incorporate prior knowledge. We show that our approach outperforms the classical Naive Bayes method and shows competitive results with SVM while having less computational complexity.

---------

Résumé: In this paper we examine the interest around a number of television series broadcast on a weekly basis. We show that through analysis of initial conversation between fans or users of dedicated web forums we can provide a description of the greatest period of interest (or peak). We then focus our attention on this peak with an ultimate goal of characterising episodes as a function of the actions and qualities of the people that take part in early conversation about this episode. We find that early interaction trends have strong similarities with the overall conversation patterns, and contain the majority of information provided by influential members of the community. This observation has important implications for the rapid generation of meta-data which may be used to during later broadcast and re-runs for description and valuation of episodes.

DOI: 10.1007/s10844-013-0235-x
Résumé: Feature-based format is the main data representation format used by machine learning algorithms. When the features do not properly describe the initial data, performance starts to degrade. Some algorithms address this problem by internally changing the representation space, but the newly-constructed features are rarely comprehensible. We seek to construct, in an unsupervised way, new features that are more appropriate for describing a given dataset and, at the same time, comprehensible for a human user. We propose two algorithms that construct the new features as conjunctions of the initial primitive features or their negations. The generated feature sets have reduced correlations between features and succeed in catching some of the hidden relations between individuals in a dataset. For example, a feature like sky ∧ ¬building ∧ panorama would be true for non-urban images and is more informative than simple features expressing the presence or the absence of an object. The notion of Pareto optimality is used to evaluate feature sets and to obtain a balance between total correlation and the complexity of the resulted feature set. Statistical hypothesis testing is used in order to automatically determine the values of the parameters used for constructing a data-dependent feature set. We experimentally show that our approaches achieve the construction of informative feature sets for multiple datasets.

---------

Résumé: La gestion des réclamations est un élément fondamental dans la relation client. C’est le cas en particulier pour la Caisse Nationale des Allocations Familiales qui veut mettre en place une politique nationale pour faciliter cette gestion. Dans cet article, nous décrivons la démarche que nous avons adoptée afin de traiter automatiquement les réclamations provenant d’allocataires de la CAF du Rhône. Les données brutes mises à notre disposition nécessitent une série importante de prétraitements, prétraitements nécessaires pour rendre les données utilisables. Une fois ces données correctement nettoyées, l’application de techniques issues de l’analyse des données et de l’apprentissage non supervisé nous permettent d’extraire à la fois une typologie des réclamations basée sur leur contenu textuel et une typologie des allocataires réclamants. Après avoir présenté ces deux typologies, nous les mettons en correspondance afin de voir comment les allocataires se distribuent selon les différents types de réclamation.

---------

Résumé: Les méthodes dédiées à l’extraction automatique de thématiques sont issues de domaines variés : linguistique computationnelle, TAL, algèbre linéaire, statistique, etc. A ces méthodes spécifiques, peuvent s’ajouter des méthodes adaptées d’autres domaines, notamment de l’apprentissage automatique non supervisé. Les résultats produits par l’ensemble de ces méthodes prennent des formes hétérogènes : partitions de documents, distributions de probabilités sur les mots, matrices, etc. Cela pose clairement un problème pour les comparer de manière uniforme. Dans cet article, nous proposons une nouvelle mesure de qualité, intitulée Vraisemblance Généralisée, pour permettre une évaluation et ainsi la comparaison de différentes méthodes d’extraction de thématiques. Les résultats, obtenus sur un corpus de documents Web autour des élections présidentielles françaises de 2012, ainsi que sur le corpus Associated Press, montrent la pertinence de la mesure proposée.

Résumé: We propose an original approach for mapping the choreographic data into a new representation language adapted to data mining techniques. Our approach relies mainly on the notion of “dance tags” that we took from the NLP community by analogy with Part-of-Speech tagging. The process starts from scores described in Labanotation and produces in a fully automatic manner a high-level, comprehensive representation of the choreographic sequence. Our experiments show that we succeed in retrieving manually translated scores with an accuracy of 85% to 94%. Using this new representation of the choreographic data, one can then perform several useful tasks in an efficient manner. Among these are: music recommendation, automated detection of dance style or genre, and ultimately any task that requires a deeper understanding of the meaning of choreographic information than traditional processing can provide. In this paper, we demonstrate the usefulness of our approach with a simple example for discriminating between classical ballet, modern ballet, and folkloric dances.

Résumé: Abstract—In this paper, we propose a new time-aware dissimilarity measure that takes into account the temporal dimension. Observations that are close in the description space, but distant in time are considered as dissimilar. We also propose a method to enforce the segmentation contiguity, by introducing, in the objective function, a penalty term inspired from the Normal Distribution Function. We combine the two propositions into a novel time-driven constrained clustering algorithm, called TDCK-Means, which creates a partition of coherent clusters, both in the multidimensional space and in the temporal space. This algorithm uses soft semi-supervised constraints, to encourage adjacent observations belonging to the same entity to be assigned to the same cluster. We apply our algorithm to a Political Studies dataset in order to detect typical evolution phases. We adapt the Shannon entropy in order to measure the entity contiguity, and we show that our proposition consistently improves temporal cohesion of clusters, without any significant loss in the multidimensional variance.

---------