Search

International audience; Our communication takes place in the context of the French-German project SegCor (Segmentation of Oral Corpora, ANR-15-FRAL-0004), focusing on the segmentation of oral corpora. The general aim is the development of a method of segmentation for oral corpora that is adequate for the analyses of interactional data at different levels and for various communities of researchers.The French and German datasets consist of ten excerpts of ten minutes each for each language[3], which represent the overall data diversity in terms of situation types. The following recorded interactions have been studied: radio talks, meal preparations, reading activities with a child, service encounters, telephone calls, table talks, social meetings, school lessons and panel discussions. In our paper, we will address the relationship between these interaction types and segmentation in maximal units. More particularly, the focus will be on the composition of this kind of units for the French corpus.Several models have been proposed in previous researches and have been discussed within the SegCor project: part-of-speech tagging and chunking processes via automatic annotation (Eshkol-Taravella et al. 2014); a syntactic annotation relying on a dependency parser (Kahane et al. 2017); a macrosyntactic segmentation in illocutionary units (Benzitoun et al. 2010; Lacheret et al. 2014); the annotation of prosodic prominences and disfluencies leading to the segmentation of intonational periods (Lacheret et al. 2014); the annotation of Turn-Constructional Units (TCUs), i.e. the minimal, emergent and negotiable units through which participants build turns of talk in interaction (Sacks et al. 1974; Ochs et al. 1996; Traverso 2016).In this paper, we will focus on the segmentation of broad units, which is grounded on the macrosyntactic model (Blanche-Benveniste et al. 1990; Blanche-Benveniste 2010a, 2010b; Lacheret et al. 2014). We rely on the following maximal macrosyntactic units:Simple units, composed of one nucleus, which is defined as a minimal macrosyntactic component corresponding to an autonomous utterance, according to Blanche-Benveniste et al. (1990: 114);Complex units, composed of more than one nucleus (including pre-nuclei, post-nuclei and in-nuclei, i.e. sequences beyond government);Abandoned units, i.e. syntactically unfinished units.The segmentation has been realized on tokenized transcripts through the EXMARaLDA Partitur Editor[4]. Our main aim is to appreciate the relevance of tokens’ number per maximal unit in our representative corpora. Thus, we propose a quantitative study that is focused on token count per maximal unit in each situation type.For example, preliminary investigation has shown a higher rate of abandoned units when interactions are conflictual (e. g. panel discussion and radio talk), due to turn-taking specificities. Conversely, in expert talk, i.e. a conference realized by a speaker, abandoned units are very few because of the planned character of the talk.Relying on the composition of maximal segmentation units, our contribution discusses evidence from corpus segmentation and aims at investigating variation across different interaction types. Our approach is not in contrast to previous research in the field of corpus linguistics, see for example Biber’s multi-dimensional analyses of written and oral genres (Biber 1988) and conversational text types (Biber 2004) in English, which are based on a variety of linguistic features. This contribution offers complementary dimensions for a classification of interaction types, from a quantitative perspective. We will then explore the other segmentation levels annotated in the SegCor project on syntax, prosody and interaction to study if unit characterization depends on the type of interaction and if similar trends can be observed. Statistical analyses and graphing are performed using the R software platform.

International audience

International audience

International audience

This poster presents Atomic, a multi-level annotation tool for linguistic data. In the past, a number of annotation tools – like Synpathy , RSTTool, and MMAX2 – have been developed in linguistics. Most of them have been implemented within research projects for a specific research question or a specific type of annotation, e.g., syntactic annotations or coreference chains. Unfortunately, some of these tools have not been developed further, beyond their respective project duration. In contrast, Atomic aims at being not Yet Another Annotation Tool for Yet Another Research Question. It has been designed with a focus on adaptability, in order to future-proof it for new use cases. Atomic is powered by Salt (Zipser & Romary, 2010), a graph-based, theory-neutral, and semantic-free metamodel for linguistic data. The data abstraction via Salt and the inclusion of a generic, graph-based editor enable Atomic to handle potentially all types of annotations. Atomic was built on top of the Eclipse RCP (McAffer et al., 2010), an application framework with a sophisticated plugin technology which allows to easily extend the software for different research needs, e.g., with editors for specific annotation types. There are tried and tested annotation techniques – and editor types – for different types of annotations. In Atomic it is possible to include all of these, and make them all work on the same data model. The combination of a generic data model and a pluggable architecture, therefore, allows Atomic to outlive its original development context, as new editors and other tooling may always be added to support future theories and annotation types. To demonstrate its extensibility, Atomic ships with dedicated editors, such as a coreference editor. In order to further ensure the sustainability of the software, Atomic integrates the Pepper converter framework (Zipser et al., 2011), which enables it to read and write established linguistic formats like TigerXML, the EXMARaLDA and RST format, MMAX2, TCF, ANNIS, and many more.

Concordances have always played an important role in the analysis of language corpora, for studies inhumanities, literature, linguistics, translation and language teaching. However, very few of the available systemssupport multi-level queries against a richly-annotated, sound-aligned spoken corpus. The rapid growth in thedevelopment of spoken corpora, particularly for French, increases the need for scalable, high-performance solutions.We present the preliminary results of our project to develop a multi-level multimedia concordancer for spoken languagecorpora. We test our prototype on the PFC corpus of spoken French (1.5 million tokens, transcriptions aligned to theutterance level). Our tool allows researchers to query the corpus and produce concordances correlating severalannotation levels (part-of-speech tags, lemmas, annotation of phonological phenomena such as the liaison and schwa,etc.) while allowing for multi-modal access to the associated sound recordings and other data., Les concordanciers jouent depuis longtemps un rôle important dans l’analyse des corpus linguistiques, tout comme dans les domaines de la philologie, de la littérature, de la traduction et de l’enseignement des langues. Toutefois, il existe peu de concordanciers qui soient capables d’associer des annotations à plusieurs niveaux et synchronisées avec le signal sonore. L’essor des grands corpus de français parlé introduit une augmentation desexigences au niveau de la performance. Dans ce travail à caractère préliminaire, nous avons développé un prototype deconcordancier multi-niveaux et multimédia, que nous avons testé sur le corpus de français parlé du projet Phonologie du Français Contemporain (PFC, 1,5 million de tokens de transcription alignée au niveau de l’énoncé). L’outil permet non seulement d’enrichir les résultats des concordances grâce aux données relevant de plusieurs couches d’annotation du corpus (annotation morphosyntaxique, lemme, codage de la liaison, codage du schwa etc.), mais aussi d’élargir les modalités d’accès au corpus.

Previously, linguistic text analysis is performed manually. However, today there are new methodological approaches that allow researchers more efficient ways of working. A new methodological approach is the usage of Text Mining methods and techniques. The methodology of Text Mining comes originally from computer science and has been, inter alia, used for the purpose of frequency and co-occurrence analysis. Subject of this work is the combination of the linguistic evaluation theory and Text Mining. The linguistic evaluation theory by Sandig (1979) describes the act of evaluation as a linguistic act, where individual evaluation criteria are verbalized by linguistic expressions. Verbal evaluation is done by using an inventory of evaluative expressions, which may vary depending on text type, such as for the text type blog comment. These text types are characterized in comparison to genuine written linguistically oriented text types (e.g. journalistic texts) by specific linguistic phenomena such as standard deviation, interactive units within the meaning of Zifonun et al. (1997) or onomatopoeic expressions. They have to be taken into account in the method development. Concerning this background, the following research questions arise: What does a machine (and thus the automation) afford to meet these challenges? What evaluation-related linguistic phenomena must be considered in the automation? How do Text Mining methods and techniques be developed further to meet these challenges? The aim of this work is the development of a first theoretical-methodological approach for the automatic analysis of verbal evaluations in blog comments by Text Mining considering text type-specific linguistic phenomena. For this purpose, existing manual and automatic methods of text analysis are adapted and optimized. The methodology development is carried out using the example of blog comments based on a corpus dealing with mobile communication systems. The methodology development provides a two-stage procedure consisting of a pre- and a main study. Subject of the pre-study is the evaluation of Text Mining methods and the identification of empirical problems by manual and automatic analysis of blog comments; the evaluation of automatic analysis tools is done with reference to the text analysis software PASW Modeler and MySQL. To determine the empirical problems, frequency, co-occurrence and sentiment analysis are carried out. The identified empirical, text type-related problems are classified and used as implications for the main study. Subject of the main study is to develop an approach for the identification of opinion- indicating expressions in German blog comments. The data base consists of two sub-corpora collected criteria-based. The sub-corpora are analyzed with methods and tools of corpus linguistics and Natural Language Processing, the analysis is carried out by means of linguistic multi-level annotation (in EXMARaLDA). Final result is a fine-grained, feature-based, linguistic multi-level annotation model, which is evaluated by inter-annotator agreement. Using the multi-level annotation model the sub-corpora are finally annotated by five annotators and a gold standard is derived.

Bachelor thesis "Possibilities of Error Annotation of Non-Native Speakers' Czech" compares annotation systems of selected learners corpora from the perspective of error annotation. For the comparison, two learner corpora were chosen - Czech CZESL and German FALKO. Both corporas use stand-off multi-level annotation model. The paper is divided into two parts: theoretical and practical. In the theoretical part there is an in-depth description of both of selected corporas and their annotation models. The practical part presents annotation of pupil text processed in annotation models of both corporas. The aim of this paper is to highlight possible strengths and weaknesses of selected annotation formats.

Colloque avec actes et comité de lecture. internationale.; International audience; Les corpus français librement accessibles annotés à d'autres niveaux linguistiques que morpho-syntaxique sont insuffisants à la fois quantitativement et qualitativement. Partant de ce constat, la FREEBANK définit une base de corpus du français annotés à plusieurs niveaux (structurel, morphologique, syntaxique, coréférentiel) et à différents degrés de finesse linguistique qui soit libre d'accès, codée selon des schémas normalisés, intégrant des ressources existantes et ouverte à l'enrichissement progressif.

International audience; Cette présentation décrira l'approche méthodologique et les différents problèmes posés par la segmentation multi-niveaux de corpus oraux dans le projet franco-allemand SegCor (SEGmentation de CORpus oraux, ANR-15-FRAL-0004), qui réunit l'IDS (Institut für Deutsche Sprache) de Mannheim, le LLL (Laboratoire Ligérien de Linguistique) d'Orléans et le laboratoire ICAR de Lyon. Ce projet a pour but de développer une méthode de segmentation utilisable pour l'analyse de la parole-en-interaction dans différentes perspectives et pour différentes communautés de chercheurs.À partir des banques de données ESLO et CLAPI pour le français et FOLK pour l'allemand, un corpus pilote de 10 extraits de 10 minutes d'interactions de nature différente a été constitué pour établir un jeu de données comparable dans les deux langues. Différentes approches de segmentation décrites en syntaxe, en prosodie, en linguistique interactionnelle et en linguistique de corpus ont alors été choisies et mises en œuvre par les trois équipes de recherche.Au sein de notre laboratoire, les niveaux de segmentation macrosyntaxique et microsyntaxique ont fait l'objet d'un travail comprenant plusieurs phases. En nous appuyant sur les projets ANR Rhapsodie et ORFEO, nous avons envisagé des phases préliminaires d'annotation, l'évaluation de l'accord inter-annotateur, l'examen de cas critiques et de nouvelles phases d'annotation. Pour ces deux niveaux, le travail a été effectué sur le corpus pilote et l'implémentation du codage est en cours de formalisation à l'aide de protocoles spécifiques (guidelines). Pour l'annotation interactionnelle en unités de construction de tour (Turn Constructional Units), que nous développerons plus en détail, un projet exploratoire porte actuellement sur la discussion des différentes approches et des critères qui permettent d'objectiver le repérage de telles unités en français et en allemand.L'annotation prosodique en proéminences intonatives a fait l'objet d'une expérience de codage manuel et automatique sur la totalité du corpus pilote. Elle nous a amené à proposer quelques modifications au protocole Rhapsodie.Le niveau d'annotation du chunking, réalisé par le LLL, a bénéficié d'un travail d'apprentissage automatique sur corpus, avec correction manuelle.Dans cette présentation, nous ne nous pencherons pas seulement sur la description des différents niveaux de segmentation et des jeux d'étiquettes, mais aussi sur les questions de conception de procédures d'annotation et de comparaison de résultats, discutées entre les trois partenaires du projet. Nous aborderons également les enjeux méthodologiques et les contraintes techniques posées par certains logiciels (EXMARaLDA, PRAAT).