N-gramme | EPFL Graph Search

Un n-gramme est une sous-séquence de n éléments construite à partir d'une séquence donnée. L'idée semble provenir des travaux de Claude Shannon en théorie de l'information. Son idée était que, à partir d'une séquence de lettres donnée (par exemple « par exemple ») il est possible d'obtenir la fonction de vraisemblance de l'apparition de la lettre suivante. À partir d'un corpus d'apprentissage, il est facile de construire une distribution de probabilité pour la prochaine lettre avec un historique de taille n. Cette modélisation correspond en fait à un modèle de Markov d'ordre n où seules les n dernières observations sont utilisées pour la prédiction de la lettre suivante. Ainsi un bigramme est un modèle de Markov d'ordre 2. À titre d'exemple, le bi-gramme le plus fréquent de la langue française est « de », comme dans l'article « de », mais aussi comme dans les mots « demain », « monde » ou « moderne ». En traitement du langage naturel il est f

Analyse multi-échelle de n-grammes sur 200 années d'archives de presse

Vincent Christian Buntinx

The recent availability of large corpora of digitized texts over several centuries opens the way to new forms of studies on the evolution of languages. In this thesis, we study a corpus of 4 million press articles covering a period of 200 years. The thesis tries to measure the evolution of written French on this period at the level of words and expressions, but also in a more global way by attempting to define integrated measures of linguistic evolution. The methodological choice is to introduce a minimum of linguistic hypotheses in this study by developing new measures around the simple notion of n-gram, a sequence of n consecutive words. The thesis explores on this basis the potential of already known concepts as temporal frequency profiles and their diachronic correlations, but also introduces new abstractions such as the notion of resilient linguistic kernel or the decomposition of profiles into solidified expressions according to simple statistical models. Through the use of distributed computational techniques, it develops methods to test the relevance of these concepts on a large amount of textual data and thus allows to propose a virtual observatory of the diachronic evolutions associated with a given corpus. On this basis, the thesis explores more precisely the multi-scale dimension of linguistic phenomena by considering how standardized measures evolve when applied to increasingly long n-grams. The discrete and continuous scale from the isolated entities (n=1) to the increasingly complex and structured expressions (1 < n < 10) offers a transversal axis of study to the classical differentiations that ordinarily structure linguistics: syntax, semantics, pragmatics, and so on. The thesis explores the quantitative and qualitative diversity of phenomena at these different scales of language and develops a novel approach by proposing multi-scale measurements and formalizations, with the aim of characterizing more fundamental structural aspects of the studied phenomena.

EPFL2017

Benign synthesis of quinolinecarboxamide ligands, H(2)bqbenzo and H(2)bqb and their Pd(II) complexes: X-ray crystal structure, electrochemical and antibacterial studies

Maryam Bagheri, Farzaneh Fadaei Tirani

Two carboxamide ligands, H(2)bqbenzo {3,4-bis(2-quinolinecarboxamido)benzophenone} and H(2)bqb {N,N-bis[(2-quinolinecarboxamide)-1,2-benzene]}, have been prepared using tetrabutylammonium bromide as an environmentally benign reaction medium. Two new Pd(II) complexes, [Pd-II(bqbenzo)] (1) and [Pd-II(bqb)] (2), have been synthesized, characterized, and their structures determined by single crystal X-ray diffraction. The di-anionic ligands, bqbenzo(2-) and bqb(2-), are coordinated via two N-amide atoms and the nitrogens of the two quinoline rings, with Pd-N-amide

Taylor & Francis Ltd2017

Semantic Vector Machines

Vincent Etter

We first present our work in machine translation, during which we used aligned sentences to train a neural network to embed n-grams of different languages into an d-dimensional space, such that n-grams that are the translation of each other are close with respect to some metric. Good n-grams to n-grams translation results were achieved, but full sentences translation is still problematic. We realized that learning semantics of sentences and documents was the key for solving a lot of natural language processing problems, and thus moved to the second part of our work: sentence compression. We introduce a flexible neural network architecture for learning embeddings of words and sentences that extract their semantics, propose an efficient implementation in the Torch framework and present embedding results comparable to the ones obtained with classical neural language models, while being more powerful.

2009

Analyse multi-échelle de n-grammes sur 200 années d'archives de presse

Vincent Christian Buntinx

EPFL2017