The Quaero program is an international project promoting research and industrial innovation on technologies for automatic analysis and classification of multimedia and multilingual documents. Within the program framework, research organizations and industrial partners collaborate to develop prototypes of innovating applications and services for access and usage of multimedia data. One of the topics addressed is the translation of spoken language. Each year, a project-internal evaluation is conducted by DGA to monitor the technological advances. This work describes the design and results of the 2011 evaluation campaign. The participating partners were RWTH, KIT, LIMSI and SYSTRAN. Their approaches are compared on both ASR output and reference transcripts of speech data for the translation between French and German. The results show that the developed techniques further the state of the art and improve translation quality.

Statistical Language Models (LM) are highly dependent on their training resources. This makes it not only difficult to interpret evaluation results, it also has a deteriorating effect on the use of an LM-based application. This question has already been studied by others. Considering a specific domain (text prediction in a communication aid for handicapped people) we want to address the problem from a different point of view: the influence of the language register. Considering corpora from five different registers, we want to discuss three methods to adapt a language model to its actual language resource ultimately reducing the effect of training dependency: (a) A simple cache model augmenting the probability of the n last inserted words; (b) a user dictionary, keeping every unseen word; and (c) a combined LM interpolating a base model with a dynamically updated user model. Our evaluation is based on the results obtained from a text prediction system working on a trigram LM.

Les modèles markoviens de langage sont très dépendants des données d’entraînement sur lesquels ils sont appris. Cette dépendance, qui rend difficile l’interprétation des performances, a surtout un fort impact sur l’adaptation à chaque utilisateur de ces modèles. Cette question a déjà été largement étudiée par le passé. En nous appuyant sur un domaine d’application spécifique (prédiction de texte pour l’aide à la communication pour personnes handicapées), nous voudrions l’étendre à la problématique de l’influence du registre de langage. En considérant des corpus relevant de cinq genres différents, nous avons étudié la réduction de cette influence par trois modèles adaptatifs différents : (a) un modèle cache classique favorisant les n derniers mots rencontrés, (b) l’intégration au modèle d’un dictionnaire dynamique de l’utilisateur et enfin (c) un modèle de langage interpolé combinant un modèle général et un modèle utilisateur mis à jour dynamiquement au fil des saisies. Cette évaluation porte un système de prédiction de texte basé sur un modèle trigramme.

In the past decade, Latent Semantic Analysis (LSA) was used in many NLP approaches with sometimes remarkable success. However, its abilities to express semantic relatedness were not yet systematically investigated. This is the aim of our work, where LSA is applied to a general text corpus (German newspaper), and for a test vocabulary, the lexical relations between a test word and its closest neighbours are analysed. These results are compared to the results from a collocation analysis.