Contenu du sommaire : Analyse d'opinions sur internet

In this paper we offer our views on the epistemological background and the history of opinion mining. Using examples taken from current research, we show the situation of the practice in the context of information theory. We focus on the issue of identification and restitution of the opinion sender and his/her values based on text analysis. Enunciative and text theories offer some clues to enhance performance of subjectivity analysis efficiency.

The development of new services (e.g. blogs, forums…) provides facilities to express opinion on different topics. Recently new techniques known as opinion mining have emerged. They often provide statistical analyses of available opinions as an overall trend on different topics. Nevertheless, whereas it is rather easy to get such an overall trend, a more detailed analysis would highlight that cybernauts' appraisals are related to multiple criteria. For example, a film can be classified as enthusiastic because of its great scenario, disappointing because of its poor soundtrack, etc. Our objective in this paper is to automatically extract any textual segment related to a criterion (or a subset of criteria) from a critics database. Then, for any criterion, opinion-mining techniques allow computing a partial assessment with regard to the criterion from the textual segments related to it. From a short list of key words characteristic of a criterion, our approach provides the required learning database for critics classification without further human intervention. Experiments have been carried out to illustrate and validate the efficiency of our approach.

We study in this paper the sentiment classification problem using supervised classifiers corresponding to different combinations of loss and regularization functions. A key difficulty of sentiment classification compared to thematic classification is the definition and selection of relevant features. We show that classical regularization approaches fail to weight or select relevant terms for this task. We introduce a new method based on a term dependent regularization for penalizing specific families of terms. This new regularization framework enables us to obtain very good performance on a classical sentiment recognition task as well as for multi-domain sentiment classification on classical benchmarks. We provide an in depth analysis of the results and an interpretation of the model behavior. It shows that the proposed regularization method is an effective selection tool for discriminative terms in the context of sentiment classification.

Choices are often difficult to assume, especially concerning reading practices of serial literature, romances, mangas… Readership of this kind of popular romances is always depreciate, often stigmatized. We identified web forums and blogs mentioning the subject of readers defending their choice of readings, and thus revealing their feelings about their possible bad choices and bad taste. We'll see how we can analyze the data, particularly the expression of feelings with tools through methods of automatic analysis of texts.

The development of health-based social networks is part of a historical process leading to a stronger regognition of the patient's feeelings and opinions by the medical institution. These networks are often referred to as “communities”, because they host the intense socialities of users trying to overcome the loneliness caused by their diseases. But behind the word “community” lies a diversity of opinions, ideas, and feelings of belonging. Nonetheless some social networks, where users are driven by common goals and shared values, help to construct lay knowledge. We study the benefits and the drawbacks of this phenomenon, on the example of Yahoo autisme.

This paper presents the students' opinion about an experience of using the e
Portfolio as a tool reinforcing a course in a continuous training. A questionnaire is submitted to students to evaluate the students' appreciation of the ePortfolio. Once the context of this experience is presented, the questionnaire and its results are detailed.

Music emotion recognition is an industrial and academic challenge. While we observe an explosion of multimedia content, it becomes crucial to build structured sets of terms and concepts so as to help knowledge organization. In music libraries, sentiment analysis and emotion classification are emerging and remain a subject of ambitious research. In this chapter, we will do an overview of the state-of-the-art over the application context and the quantitative and qualitative assessment of feelings through the intersubjectivity of the subject. Then we will present our experimental protocol for evaluating the intersubjectivity of the subject and the first results obtained so far.