Search

International audience; Since few years, some tools that are helping us to interpret results of deep learning have appeared (LIME, LSTMVIS, TDS). In this paper, we propose to go further by searching hidden information encoded in intermediate layers of deep learning thanks to a new tool. Hyperdeep allows, on the one hand, to predict the belonging of a text and to appreciate its borrowings from different styles or authors and, on the other hand, it allows to analyze, by deconvolution, the spatial and static patterns of the text in order to bring up the linguistic markers learned by the network. This new type of linguistic objects is gathered and highlighted in a graphical tool combining visualizations and hypertext. This tool is fully integrated in the Hyperbase Web platform, which offers the adequate environment and a natural starting point for any study mixing deep learning and text mining. Abstract 2 (in French) Depuis peu, les outils d'aide à l'interprétation des résultats du deep learning font leur apparition (LIME, LSTMVIS, TDS). Dans cette communication nous proposons d'aller plus loin en allant chercher l'information cachée au plus profond des couches intermédiaires du deep learning grâce à un nouvel outil. Hyperdeep permet d'une part de prédire l'appartenance d'un texte et d'en apprécier les emprunts à différents styles ou auteurs et d'autre part, par déconvolution, d'analyser les motifs spatiaux et statiques du texte afin d'en faire remonter les marqueurs linguistiques appris par le réseau. Cette information d'un genre nouveau est rassemblée et mise en valeur dans un nouvel outil mêlant visualisations graphiques et texte dynamique. Son utilisation est accompagnée d'une intégration complète dans la plateforme Hyperbase Web qui propose l'environnement adéquate et un point de départ naturel pour toute étude mêlant deep learning et statistiques du texte.; Depuis peu, les outils d'aide à l'interprétation des résultats du deep learning font leur apparition (LIME, LSTMVIS, TDS). Dans cette communication nous proposons d'aller plus loin en allant chercher l'information cachée au plus profond des couches intermédiaires du deep learning grâce à un nouvel outil. Hyperdeep permet d'une part de prédire l’appartenance d’un texte et d’en apprécier les emprunts à différents styles ou auteurs et d’autre part, par déconvolution, d'analyser les saillances du texte afin d’en faire remonter les marqueurs linguistiques appris par le réseau. Cette information d’un genre nouveau est rassemblée et mise en valeur dans un nouvel outil mêlant visualisations graphiques et texte dynamique. Son utilisation est accompagnée d’une intégration complète dans la plateforme Hyperbase Web qui propose l’environnement adéquate et un point de départ naturel pour toute étude mêlant deep learning et statistiques du texte.

This contribution compares statistical analysis and deep learning approaches to textual data. The extraction of key passages using statistics and deep learning is implemented using the Hyperbase software. An evaluation of the underlying calculations is given by using examples from two different languages—French and Latin. Our hypothesis is that deep learning is not only sensitive to word frequency but also to more complex phenomena containing linguistic features that pose problems for statistical approaches. These linguistic patterns, also known as motives Mellet and Longree (Belg J Linguist 23:161–173, 2009 [9]), are essential for highlighting key passages. If confirmed, this hypothesis would provide us with a better understanding of the deep learning black box. Moreover, it would bring new ways of understanding and interpreting texts. Thus, this paper introduces a novel approach to explore the hidden layers of a convolutional neural network, trying to explain which are the relevant linguistic features used by the network to perform the classification task. This explanation attempt is the major contribution of this work. Finally, in order to show the potential of our deep learning approach, when testing it on the two corpora (French and Latin), we compare the obtained linguistic features with those highlighted by a standard text mining technique (z-score computing).

International audience; In this paper, we propose a new strategy , called Text Deconvolution Saliency (TDS), to visualize linguistic information detected by a CNN for text classification. We extend Deconvolution Networks to text in order to present a new perspective on text analysis to the linguistic community. We empirically demonstrated the efficiency of our Text Decon-volution Saliency on corpora from three different languages: English, French, and Latin. For every tested dataset, our Text Deconvolution Saliency automatically encodes complex linguistic patterns based on co-occurrences and possibly on grammatical and syntax analysis.

International audience; We analyze a corpus of classical Latin texts, comprising various literary genres and authors. Two Correspondence Analyses (CA) are based on discrete units (used by Biber 2006). The first one represents the distances between the main works in the classical Latin corpus according to the parts of speech used in the different texts, the second according to the distribution of verb tenses and moods. The paradigmatic approach is efficient for automatically classifying the texts, but provides little new information for the linguist or philologist. We therefore assess the impact on genre characterization of taking the integration of the parts of speech (POS) and grammatical categories in syntactic structures (the syntagmatic approach) into account. However, even when the syntactic dimension is taken into consideration, this method does not really account for the sequential structure of the text’s linearity. Moreover, the choice of the syntactic structures studied depends upon the knowledge already acquired by the Latinist and their detection is always supervised. We therefore propose the new concept of motif in order to handle the different tokens of a given structure and to model them in a single pattern whose identification is based on its unified text dynamics function, disregarding surface variations. As a general pattern, the motif is able to characterize a genre; but its different realizations or tokens may be specific to different authors in a given genre. This claim is exemplified by a contrastive analysis of the style of two Latin historians who both lived at the close of the classical literary period, Caesar and Tacitus. In order to contribute to the discussion herein about what makes a “Grammar of Genres and Styles”, we would like to submit a methodological study based on textual analysis whose aim is to identify formal criteria for distinguishing between different discursive genres or authors’ styles and characterizing them according to their linguistic properties and textual dynamics1. In our previous work, we have used methods relying not only on a paradigmatic, quantitative analysis but also on syntagmatic approaches: sequences (Longree and Luong 2003, 2005), text segmentations (Longree, Luong, and Mellet 2004, 2006; Longree and Mellet 2007), neighbourhoods (Mellet and Barthelemy, 2007; Luong, Julliard, Mellet and Longree, 2007; Barthelemy, Longree, Luong, and Mellet 2009) and bursts (Longree, Luong, and Mellet 2008; Longree and Mellet 2016). This work has led to a theoretical proposal to consider the text as a topological space and to introduce a new analytical unit that we call the “motif” (Longree, Luong and Mellet 2008; Mellet and Longree 2009, 2012; Longree and Mellet 2013, 2014). With this methodological background in mind, we would like to assess here the benefits and limitations of both approaches - paradigmatic and syntagmatic - in the characterization of textual genres and author’s styles.

no abstract

International audience; Relying on the analysis of a Latin historical corpus, our research aims to study the markers structuring literary texts in general, and focuses on methods which, by extension, should be valid for any text of some length. Our basic assumption is the following: such texts include complex multilevel structures (i.e. those calling upon lexis, semantics, morphology, syntax...) which function as heterogeneity indicators (progression to a new episode, focalisation on a new point of view, insertion of reported speech, etc.). Additionally, the recurrence of these structures is a factor in textual cohesion. Under certain conditions, they function as topological ‘motifs’ marking the linear progression of the text and ensuring textual unity. We are developing new methods to detect and analyse the distributions of such ‘motifs’ and to support structural comparisons with the objective of contrastive corpus studies (contrasts between genres, authorial styles, etc.). Our methods are based on mathematical models (neighbourhoods, bursts) and combine a qualitative approach with a sequential quantitative analysis in order to comprehend language in a linear fashion.

International audience; Neural network models describe semantic priming effects by way of mechanisms of activation of neuron coding for the words that rely strongly on synaptic efficacies between pairs of neurons. Biologically inspired Hebbian learning defines efficacy values as a function of the activity of pre- and post-synaptic neurons only. It generates only pair associations between words in the semantic network. However, the statistical analysis of large text databases points to the frequent occurrence not only of pairs of words (e.g., “the way”) but also of patterns of more than two words (e.g., “by the way”). The learning of these frequent patterns of words is not reducible to associations between pairs of words but must take into account the higher level of coding of three-word patterns. The processing and learning of pattern of words challenges classical Hebbian learning algorithms used in biologically inspired models of priming. The aim of the present study was to test the effects of patterns on the semantic processing of words and investigates how an inter-synaptic learning algorithm succeeds at reproducing the experimental data.The experiment manipulates the frequency of occurrence of patterns of three words in a multiple-paradigm protocol. Results show for the first time that target words benefit more priming when embedded in a pattern with the two primes than when only associated with each prime in a pair. A biologically inspired, inter-synaptic learning algorithm is tested that potentiates synapses as a function of the activation of more than two pre- and post-synaptic neurons. Simulations show that the network can learn patterns of three words to reproduce the experimental results

A method initially developed to carry out automatic comparisons and classifications of Latin historians texts shows that some well-ordered and characteristic structures, – some syntactical motifs in particular –, can perfectly distinguish a text from the others texts making up the corpus. This paper will establish on which factors the relevance of such a reduction method can be grounded in order to give an account of affinities and distances between texts: recurrence of the chosen motifs, exhaustiveness of the counting, respect of the text linearity both on the local level and on the global level. This will lead to a reflection on the notions of texture and textuality, understood through the properties of linearity and reticularity.

International audience; A new and recent approach grounds phraseology not only on idiomaticity, but also on criteria such as frequency, recurrence, and memorization. In this context, this paper aims to introduce the "motif" as a new phraseological concept. While this proposal has the disadvantage of adding yet another term to an already rich and complex terminology, we will try to show that it nevertheless has rewarding advantages:-it promises to allow a discussion of the various acknowledged types of phraseological units, as well as an evaluation of the possibility of defining an integrative concept which could be able to subsume some of them, without denying the relevance of long-established oppositions;-it can widen the field of phraseology to incorporate discourse analysis by highlighting important but usually underestimated features of phraseological entities, that is to say the structuring and characterizing functions of certain text patterns.First, we will define the notion of "motif" from a formal and from a functional point of view; we will then evaluate the differences and similarities between motifs and the main types of phraseological units that are usually acknowledged by linguists. Finally, we will illustrate, with some basic Latin examples, the structuring and characterising functions of some motifs and of their discursive variations.; L’objectif de cet article est de proposer, dans le cadre d’une conception récente de la phraséologie fondée sur des critères de fréquence, de récurrence et de mémorisation, et non pas du seul figement, l’intégration d’un nouveau concept – celui de motif – à l’ensemble conceptuel et terminologique de la discipline. On montre que cette proposition, qui a certes l’inconvénient d’alourdir une terminologie déjà riche et complexe, a aussi deux avantages :-le premier est de permettre une réflexion sur les différents types d’unités phraséologiques reconnues et de suggérer la possibilité d’un concept intégratif qui, sans nier la pertinence des oppositions établies, pourrait être à même d’en subsumer certaines ;-le second est d’ouvrir le champ de la phraséologie à l’analyse de discours en mettant en exergue une des fonctionnalités importantes de la création phraséologique jusqu’ici relativement sous-estimée : en effet, dans la définition que nous en proposons, nous considérons que certains motifs peuvent jouer un rôle fondamental d’agencement discursif propre à structurer les textes et à caractériser certains usages. Ainsi, après une définition de la notion de motif sur le plan formel et sur le plan fonctionnel, on confronte le motif avec les principales unités phraséologiques traditionnellement reconnues pour en montrer les proximités et les différences. Puis, à partir d’exemples empruntés au latin, on illustre la fonction structurante et la fonction caractérisante des motifs et de leurs différentes instanciations en discours.

Cet article essaye d’évaluer l’impact de la lemmatisation, ou, inversement, de la flexion casuelle sur les réseaux cooccurrentiels d’un mot-pôle en latin. Conjointement, il exploite cet impact pour approfondir l’examen de l’asymétrie des relations de cooccurrence.Nos précédents articles méthodologiques, consacrés notamment à l’asymétrie de la cooccurrence, reposaient en effet sur le dénombrement des cooccurrents d’un mot-pôle considéré et décompté sous sa forme de lemme. Or, si la cooccurrence est bien la « forme minimale du contexte » [Mayaffre 2008] qui contribue à construire le sens textuel, la question se pose alors de la pertinence de l’abstraction opérée par la lemmatisation. Les différentes formes fléchies d’un mot en latin sont porteuses non seulement des catégories du nombre (et, pour les adjectifs et les pronoms, du genre), mais aussi de marques casuelles étroitement reliées à la fonction syntaxique et l’ensemble est susceptible d’avoir une influence sur le contexte immédiat du mot, y compris dans sa dimension thématique.Nous montrons donc dans un premier temps que la forme casuelle d’un mot latin détermine pour une bonne part la liste de ses cooccurrents spécifiques. Nous exploitons ensuite ce constat pour confirmer, au moyen de deux tests différents qui utilisent le réseau cooccurrentiel des diverses formes fléchies d’un même lemme, une hypothèse précédemment avancée [Luong et al. 2010] selon laquelle l’asymétrie des cooccurrences est, pour partie au moins, une image de la double insertion d’une forme, d’une part, dans un paradigme lexico-sémantique stabilisée en langue, d’autre part, dans un réseau de relations syntagmatiques actualisées en discours. Cooccurrential Asymmetry and ContextualizationThis papers aims to evaluate the impact of the lemmatization, or, on the other way round, of the casual inflection on co-occurrence networks organized around a word-pole in Latin. At the same time, it will provide new insights into the asymmetrical relations of co-occurrence.Our previous papers on this subject took only into account co-occurrence networks of lemmas. Now, if the co-occurrence indeed is the “minimal form of the context” [Mayaffre 2008] which contributes to build textual meaning, one can wonder if the abstraction operated by the lemmatization is appropriate in this case. The various inflected forms of a Latin word signify not only the number (and, for adjectives and pronouns the gender), but also the case which is closely connected to the syntactic function. This could have an influence on the immediate context of the word, regarding also its thematic dimension.At first, we show that the case of a Latin word conditions largely the list of its specific collocates. By leaning on this observation and on two different tests based on the study of co-occurrence networks of the various inflected forms of the same lemma, we confirm our previous hypothesis [Luong et al. 2010]: the asymmetry of the co-occurrences reflects, at least partially, the double insertion of a word, on one hand, in a lexico-semantic paradigm stabilized in language, on the other hand, in a network of syntagmatic relations realized in speech.

Nous exposons ici différentes méthodes de classification automatique des textes littéraires et nous en comparons les performances, notamment en ce qui concerne leur aptitude à traduire les structurations génériques du corpus. Nous montrons qu’une approche topologique des textes, qui prend en compte leur linéarité fondamentale, c’est-à-dire l’ordre macro- et micro-structurel de leurs différentes unités constitutives, permet d’obtenir de meilleurs résultats classificatoires que les méthodes traditionnelles qui tendent à négliger cette structure linéaire. In this paper, we present different methods of automatic classification applied to a corpus of literary texts and we compare their different results; in particular we evaluate how each of them is suitable for exhibiting the generic classification of the corpus. We demonstrate that a topological approach of the texts which takes into account their linearity, i.e. the order of their micro- and macro-structures, results in better clustering than traditional quantitative methods which leave generally out of count this linear structure.

International audience; The model described here relies on the key concepts of topology, i.e. neighbourhood and equivalence of shape. A linguistic object L is studied in text T by means of one or several local questions Q. The set of successive local answers is processed so as to provide a global function characterizing the textual space under scrutiny. We begin with short sequences of tenses to illustrate the way in which to explore originally Emile Benveniste's concepts of history and discourse . We then supply life-size examples of other objects selected for their heuristic value. We go on to demonstrate the model at work on the distribution of strings of finite (F) and non-finite (n) verbal forms in the LOB Corpus of English. A topological chart is produced as the synthetic image mirroring the locations of the relevant linguistic entities throughout the text. All the individual strings concatenating any number of F and n are classified in a table. Alternatively, individual full-text strings can be extracted. We then proceed to refine the notion of lexical distribution in "rafales" in a lemmatized corpus of Latin texts, the purpose being to test the stability of the distributions in individual texts of selected verbs and assess whether a verb's behaviour is related to its semantic status. The final section is devoted to other Latin texts. The use of segments of equal length makes it possible to draw up the narrative profile of each author as revealed by his handling of tenses in main clauses.

SummaryThe study of tense variation in Latin texts can be renewed by using new devices, such as the lemmatized LASLA files and the Estela and Hyperbase programs (CNRS - UMR 6039 “Bases, Corpus et Langage”). Tense variation can therefore be related more precisely to various factors: writer, time period or literary genre. This will be exemplified by studying a corpus of classical historical texts, from Caesar to Suetonius. This paper will especially investigate the relations between tense variation and text sub-types, i.e. annalistic history, biography, military commentaries...

Le calcul de distance entre les textes repose souvent sur l'étude de fréquences ; ainsi, les diverses œuvres des historiens latins peuvent être caractérisées par des dénombrements de catégories grammaticales, comme, par exemple, les temps et modes verbaux ; nous proposons ici une autre méthode prenant en compte la répartition des formes verbales le long de l'axe syntagmatique et rendant compte des effets de succession, – enchaînements et ruptures – , au sein de l'ensemble ordonné que constitue le texte. Des exemples mettront en évidence l'intérêt de la démarche quand il s'agit d'évaluer le poids des divers facteurs à l'origine des écarts entre textes : époque, auteur, genre littéraire… Verbal tenses and linearity of the textThe intertextual distance is often calculated by studying frequencies. For instance, the Latin historians' works have been characterized by computing frequencies of grammatical categories such as tenses and moods. The method discribed here takes into account not only frequencies of verbal forms, but also their distribution and their order in the texts, by studying the way tenses follow one another (tense sequences or ruptures). The reliability of the method to differentiate texts according to diachronic evolution, author style or literary genre will be evaluated.

no abstract

Longrée Dominique. À propos du concept d'« apposition » : les constructions rex Ancus et urbs Roma. In: L'Information Grammaticale, N. 45, 1990. pp. 8-13.

Nell'ultimo libro del romanzo di Apuleio il lessico del ritorno è riconducibile a due tipologie: i lessemi che indicano il ritorno come 'retro-trasformazione', ossia come ritorno alla dimensione fisica e sociale dell'essere umano, e quelli invece che individuano il ritorno come 'risalita' dell'anima a quella dimensione ultraterrena che costituisce, in una prospettiva platonica, la sua origine. La convergenza delle queste tematiche assume in un'importanza particolare anche ai fini dell'interpretazione complessiva dell'XI libro: la centralità del tema del ritorno, infatti, rinsalda il rapporto del finale dell'opera con il modello dell'Odissea, in particolare con il momento del ritorno di Odisseo a Itaca nel libro XIII. Poiché l'esegesi medio-platonica interpreta questo momento come un'allegoria del ritorno dell'anima alla dimensione celeste, si può supporre questa lettura si riverberi anche sul finale delle Metamorfosi apuleiane.

International audience; L’insuffisance du marquage du discours direct (verbe de dire antéposé ou en incise, proposition introductrice elliptique) pour permettre le repérage précis de l’ouverture de la séquence de discours rapporté est mise en évidence. Cette insuffisance est suppléée à des degrés divers par un certain nombre d’indices, qui servent aussi au repérage des séquences de discours direct libre. En fin de compte, les occurrences de discours direct libre n’apparaissent pas moins bien balisées, du point de vue du repérage de leur ouverture, que celles de discours direct.

International audience; Co-occurrence is an asymmetric relationship: each word forming the pair brings with it its own semantic load that comes from its own network of semantic associations. Energy and availability (Luong et al., 2010) measure this asymmetry and we suggest here that they grasp quantitatively the notion of semantic load. Re-drafting them in a probabilistic framework to make them comparable between themselves and across words, we explore their interpretative value in discourse analysis on a corpus of French political speeches. We also introduce the notion of discursive performance that articulates both notions into one, and we explore the added value it brings to availability diagnostics for the purpose of identifying the modalities of semantic contextualization of a word by a given speaker.; Les relations co-occurrentielles sont des relations asymétriques : chaque mot apporte à la paire une charge sémantique qui lui vient de son propre réseau d'associations sémantiques. Les mesures d'énergie et de disponibilité (Luong et al., 2010) quantifient cette asymétrie, et nous proposons ici qu'elles saisissent quantitativement la notion de charge sémantique. En les reformulant dans un cadre probabiliste afin de les rendre comparables entre elles et d'un mot à l'autre, nous explorons leur portée interprétative en analyse du discours sur un corpus politique français. Nous introduisons également la notion de rendement discursif pour articuler entre elles les deux notions et explorons son utilité pour nuancer le diagnostic de la disponibilité sur la modulation sémantique d'un mot par un locuteur.

International audience; Co-occurrence is an asymmetric relationship: each word forming the pair brings with it its own semantic load that comes from its own network of semantic associations. Energy and availability (Luong et al., 2010) measure this asymmetry and we suggest here that they grasp quantitatively the notion of semantic load. Re-drafting them in a probabilistic framework to make them comparable between themselves and across words, we explore their interpretative value in discourse analysis on a corpus of French political speeches. We also introduce the notion of discursive performance that articulates both notions into one, and we explore the added value it brings to availability diagnostics for the purpose of identifying the modalities of semantic contextualization of a word by a given speaker.; Les relations co-occurrentielles sont des relations asymétriques : chaque mot apporte à la paire une charge sémantique qui lui vient de son propre réseau d'associations sémantiques. Les mesures d'énergie et de disponibilité (Luong et al., 2010) quantifient cette asymétrie, et nous proposons ici qu'elles saisissent quantitativement la notion de charge sémantique. En les reformulant dans un cadre probabiliste afin de les rendre comparables entre elles et d'un mot à l'autre, nous explorons leur portée interprétative en analyse du discours sur un corpus politique français. Nous introduisons également la notion de rendement discursif pour articuler entre elles les deux notions et explorons son utilité pour nuancer le diagnostic de la disponibilité sur la modulation sémantique d'un mot par un locuteur.

International audience

no abstract

no abstract

International audience

International audience; no abstract