|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Analyzing and Reasoning on Heterogeneous Semantic Graphs
RDF Mining
Participants : Andrea Tettamanzi, Catherine Faron-Zucker, Fabien Gandon, Tran Duc Minh, Claudia d'Amato.
We carried on our investigation in an approach to RDF mining based on grammatical evolution and possibility theory, whose aim is to mine large RDF graphs by automatically generating and testing OWL 2 axioms based on the known facts. In particular, we addressed the problem of scaling up the scoring heuristics based on falsification and possibility theory we have recently proposed [36] .
Data and Knowledge Integration and Extraction
Participant : Andrea Tettamanzi.
Together with Somsack Inthasone of the National University of Laos, Nicolas Pasquier and Célia da Costa Pereira of I3S, we completed a survey on biodiversity and environment data mining [16] .
Scalable Uncertainty Management
Participant : Andrea Tettamanzi.
Within the framework of the CNR PEPS GéoIncertitude, we proposed and studied the properties of uncertain logical gates in possibilistic network, using a problem of human geography as a motivating example and testbed [28] .
Natural Language Question Answering
Participants : Andrea Tettamanzi, Elena Cabrio, Catherine Faron-Zucker, Amine Hallili.
We extended previous work on answering -relation natural language questions in the commercial domain by combining an approach to learning regular expressions based on genetic programming [21] .
Events Detection in Twitter
Participants : Amosse Edouard, Elena Cabrio, Nhan Le Thanh.
We analyze Twitter data in the objective of identifying events reported by Twitter users. Specially we have worked on two main aspects: an approach for classifying tweets as either related or not related to events and secondly we have studied an approach for disambiguating geographic entities in tweets.
We have worked on an approach for separating event-related content from the rest of Twitter messages. We have combined technics from Natural Language Processing (NLP) and Machine Learning (ML) for building a classifier model that aims at classifying tweets into two mutually exclusive classes. First of all, we apply a Named Entity Recognizer to the tweets in order to identify the occurrences of named entities and special Twitter features such as hashtags, shortened URLs or user mentions. In a second step, the named entities are replaced by their generic class in the DBpedia Ontology; we do so by using SPARQL to query the DBpedia Knowledge Base to extract the class related to each entity. Third, we use the modified content as examples to train a binary classifier. Our evaluation using different classifiers such as Naive Bayes and Long Short Term Memory have shown promising results in term of performance compared to the state of the art.
We have also worked on an approach for identifying geographic entities in Twitter. This task is challenging for two main reasons: first, a geographic term can be related to either geographic or non geographic entities (Paris can be a person or a place) and second, many geographic places might have the same name (Paris can be either the capital of France or a city in Texas). We have proposed an approach based on distant-supervision and ontology matching for identifying and disambiguate ambiguous geographic terms.