Search

A finalised digital resource of 88,000 anonymised French text messages, the 88milSMS corpus, two extracts (1,000 SMS transcoded into standardised French and 100 linguistically annotated SMS) and sociolinguistic questionnaire data were released in June 2014 for all to download via a user free-of-charge licence agreement, from the Huma-Num web service (http://88milsms.huma-num.fr, Panckhurst et al., 2014). The sud4science project (http://sud4science.org, Panckhurst et al. 2013), enabling authentic text message collection from the general public by a group of academics, is part of a vast international initiative (http://www.sms4science.org/, Fairon et al. 2006, Cougnon and Fairon, 2014, Cougnon 2015), to build a worldwide database and analyse authentic text messages in different languages. We decided to exclude full transcoding and annotation tagging in the final corpus. This is a theoretical position, since annotation is far from neutral, and is invariably linked to an interpretative framework. Owing to varying theoretical disciplinary and scientific stances, it seems that a true consensus on how to standardise the transcoding and linguistic annotation tagging does not exist (Panckhurst, 2015). Other researchers may disagree and prefer to provide both 'raw' and fully tagged corpora (Chanier et al. 2014). This theoretical position does not exclude exploring Natural Language Processing (NLP) investigation techniques, which could then be implemented in real-life applications. Examples of investigation techniques are indicated as follows: 1) Our corpus can be used to analyse current mediated electronic discourse, and help build knowledge on different SMS writing forms (Roche et al. 2015). 2) Algorithms may be used to learn from this: alignment methods for facilitating automatic transcoding have been explored (Aw et al. 2006, Beaufort et al., 2008, Guimier de Neef and Fessard, 2007, Kobus et al, 2008, Lopez et al, 2014). 3) We have devised a method for classifying 'unknown' items within text messages, which may help to automatically identify lexical 'creativity' within 88milSMS and improve electronic dictionary approaches (Lopez et al. 2015). In order to refine automatic normalisation techniques for initially non-standard texts in French, the next logical step is to compare our resource with different types of instant media (i.e. SMS, forums, tweets). Firstly, a new typology of the detected 'mistakes', based on existing typologies, will be elaborated. Secondly, automatic normalisation techniques — focusing on the most frequent errors — will be proposed. These will then be confronted with traditional automatic translation (Vilariño et al., 2012), speech recognition (Kobus et al., 2008) and spelling/grammatical checker principles (Beaufort et al., 2010). Finally, the approach should enable comparison between different types of instant media.

International audience; This article presents the system Seek&Hide, a text message processing tool developed for the sud4science LR (http://www.sud4science.org/) project. It performs the anonymisation/de-iden- ti cation of a corpus. At present, it has been used to anonymise the sud4science LR corpus of French text messages collected during the project. is is done in two phases. In the rst phase, it automatically processes over 70% of the corpus. e rest of the corpus is processed in the second phase, aided by an expert annotator via a web interface speci cally designed to simplify the task.

International audience; This paper presents two anonymisation methods to process an SMS corpus. The first one is based on an unsupervised approach called Seek&Hide. The implemented system uses several dictionaries and rules in order to predict if a SMS needs anonymisation process. The second method is based on a supervised approach using machine learning techniques. We evaluate the two approaches and we propose a way to use them together. Only when the two methods do not agree on their prediction, will the SMS be checked by a human expert. This greatly reduces the cost of anonymising the corpus.

September 2-6, 2013; International audience; In Human Computer Interaction, it is more and more clear that usability is not enough. In order to take into account the other criteria that may be relevant for design, G. Cockton introduced the notion of "worth" and the Worth Centered Design (WCD) framework for its operationalization. The WCD framework structures the development process and provides designers with a set of tools, including Worth Maps (WMs). Worth maps connect systems attributes to human ones, and as such represent a promising tool. However, they remain understudied and under-experimented. This paper presents the results of our experience with WMs. More precisely, it proposes the PEW (Perceived and Expected Worth) framework for worth mapping, reports findings from a study conducted with 5 experts regarding many aspects of WMs, and discusses future directions for research. Keywords: Interactive systems design, worth, Worth Maps (WMs).

International audience; Although many new methodological and modelling concepts have been proposed by the scientific community, current industries are still focusing their engineering design process on CAD model since they assume it is the starting point of many analyses with respect to product life cycle (CAM, FEA, LCA…). The paper presents the application of modelling concepts that lead the progressive justification of CAD model with respect to knowledge synthesis by least commitment. Design experts are first formalizing their knowledge that is therefore translated to form features and parameters (topology, position, orientation, dimensions…). The results show that this new design approach and models support design intents and rational, but the generated CAD model is not fully justified. That drives to many conclusions: CAD model is many often non-100% rational by designers’ knowledge, design solution space is therefore larger than the one modelled in CAD software and could be used to foster innovation.

International audience; There has been increasing interest in the artificial intelligence community for influencer detection in recent years for its utility in singling out pertinent users within a large network of social media users. This could be useful, for example in commercial campaigns, to promote a product or a brand to a relevant target set of users. This task is performed either by analysing the graph-based representation of user interactions in a social network or by measuring the impact of the linguistic content of user messages in online discussions. We performed independent studies for each of these methods in the present paper with a hybridisation perspective. In the first study, we extract structural information to highlight influence among interaction networks. In the second, we identify linguistic features of influential behaviours. We then compute a score of user influence using centrality measures with the structural information for the former and a machine learning approach based on the relevant linguistic features for the latter.

Introduction The electronic health record (EHR) is a very important potential source of data for various areas, such as medical decision support tools, evidence-based medicine or epidemiological surveillance. Much of this data is available in text format. Methods of natural language processing can be used to perform data mining and facilitate interpretation. The purpose of this project was to develop a generic semantic solution for extracting and structuring medical data for epidemiological analyses or for medical decision-support. The solution was developed with the objective of making it as independent as possible from the field of medical application in order to allow any new user to write his or her own expert rules regardless of their area of medical expertise. Material and methods SYNODOS offers a modular architecture that makes a clear distinction between the linguistic rules and the medical expert rules. Different modules have been developed or adapted for this purpose: an interface between the multi-terminology server and semantic analyzer during the extraction phase, linguistic rules to extract temporal expressions, expert rules adapted to two areas of application (nosocomial infections, cancer), an interface between the engine and the linguistic knowledge base. Results Modular integrations were performed consecutively. The multi-terminology extractor and semantic analyzer were first interfaced during the extraction phase. Output of this data processing was then integrated into a knowledge base. A user interface to access documents and write business rules was developed. Expert rules for the detection of nosocomial infections and for the evaluation of colon cancer management have been developed. It was necessary to develop an additional module the need for which had not been identified during the drafting of the protocol. This module aims to structure the output of the data processing described above, according to the patient's care pathway. This module is based on the writing of medical expert rules. Evaluation indicators were obtained at different stages of the process (terminology extraction, semantic relations, data structuring, detection of events of interest). Discussion This project helped to highlight the value of combining different technologies (natural language processing, terminology, expert systems integration) to allow for the use of unstructured data in epidemiology. However, the need to develop an additional module of expert rules did not allow a complete and operational solution. Furthermore the multi-terminology extractor (ECMT V2) response time is too long (6 min per report). A change in technology was envisaged at the end of the project to reduce this time. Conclusions The originality of the SYNODOS project is the development of a single solution that integrates different technologies needed for the production of epidemiological indicators in the context of hospital activity. The project results confirm the interest but certain technological obstacles concerning the processing time need to be resolved in order to render the solution operational in a hospital environment.

International audience; In the context of large-scale problems, traditional multiclass classification approaches have to deal with class imbalancement and complexity issues which make them inoperative in some extreme cases. In this paper we study a transformation that reduces the initial multiclass classification of examples into a binary classification of pairs of examples and classes. We present generalization error bounds that exhibit the interdependency between the pairs of examples and which recover known results on binary classification with i.i.d. data. We show the efficiency of the deduced algorithm compared to state-of-the-art multiclass classification strategies on two large-scale document collections especially in the interesting case where the number of classes becomes very large.

International audience; Hyper-parameter tuning is a resource-intensive task when optimizing classification models. The commonly used k-fold cross validation can become intractable in large scale settings when a classifier has to learn billions of parameters. At the same time, in real-world, one often encounters multi-class classification scenarios with only a few labeled examples; model selection approaches often offer little improvement in such cases and the default values of learners are used. We propose bounds for classification on accuracy and macro measures (precision, recall, F1) that motivate efficient schemes for model selection and can benefit from the existence of unlabeled data. We demonstrate the advantages of those schemes by comparing them with k-fold cross validation and hold-out estimation in the setting of large scale classification.

International audience; In this paper we propose a simple yet effective method for sparsifying a posteriori linear models for large-scale text classification. The objective is to maintain high performance while reducing the prediction time by producing very sparse models. This is especially important in real-case scenarios where one deploys predictive models in several machines across the network and constraints apply on the prediction time. We empirically evaluate the proposed approach in a large collection of documents from the Large-Scale Hierarchical Text Classification Challenge. The comparison with a feature selection method and LASSO regularization shows that we achieve to obtain a sparse representation improving in the same time the classification performance.

International audience; In 2011, 6 academics gathered over 90,000 authentic text messages in French from the general public, in compliance with French law. The SMS ‘donors’ were also invited to fill out a sociolinguistic questionnaire (http://sud4science.org, Panckhurst et al., 2013). The project is part of a vast international initiative, entitled sms4science (http://www.sms4science.org/, Fairon et al. 2006, Cougnon & Fairon, 2014, Cougnon 2015) which aims to build a worldwide database and analyse authentic text messages. After the sud4science SMS data collection, a pre-processing phase of checking and eliminating any spurious information and a three-step semi-automatic anonymisation phase were conducted (Accorsi et al. 2014, Patel et al., 2013). Two extracts were transcoded into standardised French (1,000 SMS) and annotated (100 SMS). The finalised digital resource of 88,000 anonymised French text messages, the ‘88milSMS’ corpus, the extracts, and the sociolinguistic questionnaire data are currently available for all to download, via a user free-of-charge licence agreement, from the Huma-Num web service, (http://88milsms.huma-num.fr, Panckhurst et al., 2014). Why decide to exclude full transcoding and annotation tagging phases? Transcoding ‘raw’ text messages into ‘standardised’ French means morpho-syntactic parsers and other natural language processing tools can ultimately analyse them. Checking spelling and grammar facilitates comprehension, but no supplementary information should be ‘injected’. What if a texter tries to simulate a certain form of oral French, for instance, by using an apostrophe, or through agglutination (‘j’sais’=‘je sais’, ‘chuis’=‘je suis’)? Should these items be transcoded or not? What about punctuation, often absent in text messages? Should one re-introduce this systematically? Researchers may have differing theoretical viewpoints. Another issue is tagging the corpus. After much scientific debate about previous experiences with other sms4science members, 8 tags were chosen for ‘88milSMS’: TYP(ography), MOD(ificiation), GRA(mmar), BIN(ettes, smileys/emojis), ABS(ence), LAN(guage), ORT(hography, spelling), DIV(erse). Like the previous transcoding phase, annotation is a source of theoretical disagreement. To highlight this, it may be difficult to decide which tag to use and double tagging may be necessary: Bone journé. The ‘scriptor’ may have voluntarily modified the two words (‘Bonne journée’ have a nice day) or may have lacked spelling knowledge. So should ‘MOD’ and/or ‘ORT’ be used? In another example: ‘Il es rentrer a 22h30 et jai eu ldroii au : jsui fatiguer, jai mal a la tete jvai me coucher.’ (He came home at 10.30pm and I got to hear: I’m tired, I have a headache, I’m going to bed), ‘rentrer’ (‘Il est rentré’) could be either a gramatical mistake (GRA), or the scriptor may have preferred using an ‘r’ (MOD) instead of pressing the ‘e’ to access the acute accent (on a smartphone). It is extremely difficult to provide satisfactory standardised tagging. We decided to limit the processing to two extracts. Our (rare) choice to exclude full transcoding and tagging is a theoretical position: annotation is far from neutral. It is directly linked to an interpretative framework. A true consensus on how to standardise the transcoding and annotation does not exist, owing to differing/varying theoretical, (pluri)disciplinary and scientific stances. We believe that no additional mark-up initiatives should be imposed upon researchers ; it seems more relevant to let them conduct their own annotation bearing their specific scientific questioning in mind, without being trapped within a unique theoretical framework. The ‘88milSMS’ resource will provide inspiration for many years to come. Our corpus can be used to analyse contemporary mediated electronic discourse, build knowledge on SMS writing forms (Panckhurst 2009), and let algorithms learn from this: alignment methods for facilitating automatic transcoding are currently being explored (Lopez et al. 2014), as are methods for classifying ‘unknown’ items for use in automatically identifying lexical ‘creativity’ within ‘88milSMS’ and also to improve electronic dictionary approaches. The resource also sheds light on ‘corpus-driven’ and ‘corpus-based’ approaches (Panckhurst 2013, Panckhurst et al. 2015). Xml encoding means that the resource will be eligible for long-term archiving with the CINES (https://www.cines.fr/). Perhaps, in the future, people will look back and explore these ‘snapshot’ resources and understand more about the evolution of scriptural practices and usages in the 21st century.

International audience; In this paper, we propose a method for aligning text messages (entitled AlignSMS) in order to automatically build an SMS dictionary. An extract of 100 text messages from the 88milSMS corpus (Panckhurst el al., 2013, 2014) was used as an initial test. More than 90,000 authentic text messages in French were collected from the general public by a group of academics in the south of France in the context of the sud4science project (http://www.sud4science.org). This project is itself part of a vast international SMS data collection project, entitled sms4science (http://www.sms4science.org, Fairon et al. 2006, Cougnon, 2014). After corpus collation, pre-processing and anonymisation (Accorsi et al., 2012, Patel et al., 2013), we discuss how "raw" anonymised text messages can be transcoded into normalised text messages, using a statistical alignment method. The future objective is to set up a hybrid (symbolic/statistic) approach based on both grammar rules and our statistical AlignSMS method.

International audience; A great deal of research on opinion mining and sentiment analysis has been done in specific contexts such as movie reviews, commercial evaluations, campaign speeches, etc. In this paper, we raise the issue of how appropriate these methods are for documents related to land-use planning. After highlighting limitations of existing proposals and discussing issues related to textual data, we present the method called Opiland (OPinion mIning from LAND-use planning documents) designed to semi-automatically mine opinions in specialized contexts. Experiments are conducted on a land-use planning dataset, and on three datasets related to others areas highlighting the relevance of our proposal.

In the literature, automatic summarization and automatic titling tasks are often merged. It seems that a short summary can be considered as a relevant title. But can we compare a title and a summary without having previously studied their criteria? This study aims at positioning the emergent task of automatic titling with regard to automatic summarization task. In this paper, we define a set of criteria according to the summary and to the title, and we analyze the results obtained with our method of automatic classification. This analysis enables to report real objectives of both tasks and to validate their relevance. Dans la littérature, les tâches de résumé et de titrage automatique sont souvent confondues. A priori, il semble qu’un résumé de quelques mots peut constituer un titre tout à fait pertinent. Mais peut-on comparer un titre et un résumé sans auparavant en avoir étudié leurs critères ? Cette étude a pour but de positionner l’émergente tâche de titrage automatique par rapport à celle de résumé automatique. Après avoir défini les critères attachés au résumé et au titre, nous analysons les résultats obtenus via notre méthode automatique de classification, permettant de rendre compte des objectifs réels des deux tâches et de valider leur pertinence.

International audience; Automatic titling of text documents is an essential task for several applications (automatic heading of e-mails, summarization, and so forth). This paper describes a system facilitating information retrieval in a set of textual documents by tackling the automatic titling and subtitling issue. Automatic titling here involves providing both informative and catchy titles. We thus propose two different approaches based on NLP, text mining, and Web Mining techniques. The first one (POSTIT) consists of extracting relevant noun phrases from texts as candidate titles. An original approach combining statistical criteria and noun phrase positions in the text helps in collecting informative titles and subtitles. The second approach (NOMIT) is based on various assumptions made on POSTIT and aims to generate both informative and catchy titles. Both approaches are applied to a corpus of news articles, then evaluated according to two criteria, i.e. informativeness and catchiness.