|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Communities and Social Interactions Analysis
Argumentation and Emotion Detection with Adaptive Sentiment Analysis
Participants : Vorakit Vorakitphan, Serena Villata, Elena Cabrio.
This PhD work just started in the context of the ANSWER project with Qwant search engine. One of the main objectives of the ANSWER project is to use emotion detection algorithms within text inquiries and sentiment analysis to provide powerful enhancements in the search results from Qwant search engine. The final goal is to extract effective and scalable indicators of sentiment, emotions, and argumentative relations in order to offer the users additional means to filter the results selected by the search engine. Powerful algorithms in state-of-art will be focused to define new criteria for filtering search results, i.e., the expression of a feeling in the answers found by the search engine. By doing as mentioned, textual elements to which we wish to associate a polarity will no longer be considered in their individuality but connected to each other by polarized relations to be analyzed in a higher level setting. Currently, the work progress is in the survey of state-of-the-art based on emotion detection algorithms and implementation of sentiment analysis. Then the next target, classification models with multi-label features based on emotion detection, will be deeply explored as a starting point of this research. Moreover, NLP related to emotional news content will be taken into account to build a novel dataset based on emotion annotation from news articles in sentence-level.
Cyberbullying Events Prevention
Participants : Pinar Arslan, Michele Corazza, Elena Cabrio, Serena Villata.
In the CREEP EIT project, we built an emotion detection classifier to automatically identify the emotion for user-generated texts such as Twitter and Instagram posts. The correlation analysis that we carried out to get a better understanding of the associations between emotions and cyberbullying instances unveiled that certain emotions (e.g., anger, joy) would be good indicative features to detect cyberbullying instances. Hence, our pipeline firstly reveals automatically detected emotion labels for social media texts to be used to detect cyberbullying instances. The automatically predicted emotion labels were used as one of the predictors for our cyberbullying detection classifier. As part of the project, we successfully built a classifier for offensive language in social media interactions for English, Italian and German using neural networks. This classifier was evaluated by participating in two shared tasks: Germeval (German offensive language detection) and Evalita (Italian hate speech detection). For the Germeval Challenge [29], two systems for predicting message-level offensive language in German tweets were used: one discriminates between offensive and not offensive messages, and the second performs a fine-grained classification by recognizing also classes of offense. Both systems are based on the same approach, which builds upon Recurrent Neural Networks used with the following features: word embeddings, emoji embeddings and social-network specific features. The model combines word-level information and tweet-level information to perform the classification tasks. Our best performing model ranked 7th out of 51 submitted runs on the binary classification task, 5th out of 25 for the fine-grained classification task. For the Evalita Challenge shared tasks [28], our submissions were based on three separate classes of models: a model using a recurrent layer, an ngram-based neural network and a LinearSVC. For the Facebook task and the two cross-domain tasks we used the recurrent model and obtained promising results, especially in the cross-domain setting. For Twitter, we used an ngram-based neural network and the Linear SVC-based model. Our system ranked 1st in the Facebook to Twitter dataset, 2nd in the Twitter to Facebook dataset, 3rd in the Facebook dataset and 4th on the Twitter dataset.
Modeling of a Social Network of Service Providers
Participants : Molka Dhouib, Catherine Faron Zucker, Andrea Tettamanzi.
In the framework of a collaborative project with Silex France company and the CIFRE PhD thesis of Molka Dhouib, our aim is to model the social network of service providers and companies registered in the software as a service sourcing tool developed by Silex for the recommendation of the service providers that are best suited to meet the service requests expressed by companies. Our aim is to automate the matching of service requests and offers by reasoning on the social network of service providers and companies. We developed an automatic categorization of companies, service requests and service offers based on their textual descriptions. We conducted some experiments using state-of-the-art supervised Machine Learning techniques to classify Silex textual data into predefined categories, and to choose the best vector representations of the textual descriptions of service offers and requests in the Silex platform, and the best Machine Learning algorithm. This work has been presented at the French conference on applications of Artificial Intelligence APIA2018 [31].