|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Communities and social interactions analysis
Community Detection and Interest Labeling
Participants : Zide Meng, Fabien Gandon, Catherine Faron-Zucker.
Topic Modeling Based Overlapping Community Detection
Based on previous work, we conducted more experiments to evaluate the effectiveness and efficiency of the proposed tag tree based method. We used perplexity score to evaluate the performance of topic extraction. We got consistent performance when applying the model on a Flickr dataset. This work has been published in IEEE/WIC/ACM Web Intelligence 2015 [32] and Social Network Analysis and Mining Journal [13] .
Temporal Analysis of User and Topic
By jointly modeling topic, expertise, time and activity, we were able to retrieve many meaningful latent information from the user generated contents. We proposed a method to track the dynamics of topics and users. It can also track the dynamics with a specific granularity of time level such as, yearly, monthly, daily and hourly. Besides, the model can overcome a comparison problem of LDA based model by modeling the reverse distribution.
Topic labeling
The output of topic model is normally a bag of words. Each topic consists of closely related words. An interesting question is to assign one or more topic label to this set in order to indicate the general meaning of a bag of words. By integrating the original dataset with linked open data sources, we are now planning to propose a generic method to automatically label the detected topics.
Semantic Modeling of Social, Spatiotemporal and Dedicated Networks
Participants : Amel Ben Othmane, Nhan Le Thanh, Michel Buffa, Andrea Tettamanzi, Serena Villata.
We have been working on modeling a multi-agent based recommender system. The aim of such system is to recommend a list of activities (plans) according to user preferences in order to achieve a goal. For this purpose, we propose a multi-context framework based on the well-known agent Belief-Desire-Intention (BDI) architecture [58] . First, we extend the BDI model with additional contexts in order to handle sociality. Second, we use a possibilistic approach based on the work of Da Costa Pereira & Tettamanzi [55] , to reason about beliefs, desires, goals and intentions. Further, we use ontologies to represent and reason about plans and intentions. The proposed framework is detailed in a long paper that will be presented in the 8th International Conference on Agents and Artificial Intelligence in 2016 [18] .
Collaborative Software Development Platforms
Participant : Isabelle Mirbel.
The collaborative nature of software development helped in the emergence of several online collaborative software development platforms (CSDPs). These platforms enable distributed teams of contributors to participate in the development of the various hosted projects. In such a context, the identification of relevant contributors is very important for handling efficiently the abundant requirements. However, this can be really challenging because of the fairly large number of involved contributors, especially in some distinguished projects. Moreover, the contributor profiles on a CSDP are often inadequately informative, which makes them an unqualified resource for learning about the contributors.
In this context, we proposed to identify contributors by their reputation on a CSDP. Our approach calculates reputation scores using a belief calculus, called subjective logic, according to contributors' performed roles. Knowing the reputation of anonymous contributors would enable project members to reduce the uncertainty in their future interactions with them. Moreover, we use concept lattices to classify contributors by their reputation scores, which enable us to have a comparable view on the considered contributors. Consequently, we can produce a roadmap to examine new requirements thus supporting their effective communication and prioritization.
Logical Foundations of Cognitive Agents
Participants : Andrea Tettamanzi, Serena Villata.
Together with Célia da Costa Pereira (I3S, UNS), we have continued an investigation about the issue of trust in multi-agent systems, and we proposed a computational model of trust based on the content of messages and on the characteristics of their sources [27] .
Combining Argumentation Theory and Normative Reasoning with Natural Language Processing
Participants : Serena Villata, Elena Cabrio.
Together with Cristian Cardellino and Laura Alonso Alemany from the University of Cordoba (Argentina), we applied different Active Learning strategies to Information Extraction from licenses in English, with highly repetitive text, few annotated or unannotated examples available, and very fine precision needed. We showed that the most popular approach to active learning, i.e., uncertainty sampling for instance selection, does not provide a good performance in this setting. We showed that we can obtain a similar effect to that of density-based methods using uncertainty sampling, by just reversing the ranking criterion, and choosing the most certain instead of the most uncertain instances. The results of this research have been published at the CICLing [24] and JURIX [23] international conferences.
In another work, together with Alessio Palmero Aprosio (FBK Trento, Italy), we have worked on an extension of QAKiS, the system for open domain Question Answering over linked data, that allows to query DBpedia multilingual chapters. Such chapters can contain different information with respect to the english version, e.g. they provide more specificity on certain topics, or fill information gaps. In particular, we have introduced and evaluated the RADAR 2.0 framework for information reconciliation over language-specific DBpedia chapters. The framework is composed of three main modules: a module computing the confidence score of the sources depending either on the length of the related Wikipedia page or on the geographical characterization of the queried entity, a module retrieving the relations holding among the elements of the results set, and finally a module computing the reliability degree of such elements depending on the confidence assigned to the sources and the relations among them. This third module is based on bipolar argumentation theory to return the acceptability degrees. A demo of the RADAR framework is available online (http://qakis.org/qakis2 ). This contribution has been submitted to the Semantic Web Journal and is under review.
Moreover, we have proposed the BEGincs (BEG-Inconsistencies) framework, which translates a bipolar entailment graph into an argumentation graph. It then provides to the annotators sets of arguments that are supposed to be consistent. If it is not the case, the Textual Entailment system wrongly assigned some relations. Moving from single pairs to an overall graph allows for the detection of inconsistencies otherwise undiscovered. BEGincs does not identify the precise relation causing the inconsistency, but provides annotators with the consistent arguments sets, they are supported in narrowing the causes of inconsistency. The results of this research have been published at the CLIC conference [43] .
Argumentation theory and its applications
Participants : Elena Cabrio, Serena Villata, Fabien Gandon, Andrea Tettamanzi.
Together with Celia da Costa Pereira (UNS), we have proposed a framework to measure the acceptability of an information in a multiagent system, according to (i) the agent's goals and the information source's goals, (ii) the credibility, for the agent, of the incoming information and (iii) the agent's beliefs (or perceptions) about the context (or situation) in which it operates. The results of this research have been published at the AAMAS international conference [27] .
Moreover, together with Sahbi Benlamine, Maher Chaouachi and Claude Frasson (U. of Montreal) we have presented an empirical evaluation of the relationship between the argumentative structures of human debates and the emotions felt by the debate participants. Argumentation is often seen as a mechanism to support different forms of reasoning such that decision-making and persuasion, but all these approaches assume a purely rational behavior of the involved actors. However, humans are proved to behave differently, mixing rational and emotional attitudes to guide their actions, and it has been claimed that there exists a strong connection between the argumentation process and the emotions felt by people involved in such process. We assess this claim by means of an experiment: during several debates people's argumentation in plain English is connected and compared to the emotions automatically detected from the participants. The results of this research have been published at the IJCAI international conference [19] , and submitted to the Cognitive Science journal (under review).
Natural Language Argumentation on Twitter
Participants : Tom Bosc, Elena Cabrio, Serena Villata.
A great amount of textual data is published on social media every day. For example, there are about 500 million new tweets per day on Twitter. These data reflect the opinion and thoughts of a large population and are thus potentially useful to decision-makers and marketers, among others. But processing them is challenging because of their large quantity as well as their noisiness : poor quality of writing, redundancy, presence of advertisement, etc.
The goal of this project is to build a pipeline to automatically analyze messages exchanged on Twitter and build informative and synthetic views. We study tweets under the angle of argumentation theory. First of all, the algorithm filters in argumentative tweets. Then, it describes how tweets relate to one another : tweets may support or attack other tweets, or be neutral. Finally, a visualisation of the interactions between tweets is produced. Individual parts of the pipeline are machine learning models that are trained using datasets that are crafted specially for the project. Importantly, datasets span several domains (politics, society topics, product announcements) to ensure that the approach is generic enough and will generalize to unseen topics.