Author Profiling en Social Media: Identificación de Edad, Sexo y Variedad del Lenguaje

[EN] The possibility of knowing people traits on the basis of what they write is a field of growing interest named author profiling. To infer a user's gender, age, native language or personality traits, simply by analysing her texts, opens a wide range of possibilities from the point of vie...

ver descrição completa

Detalhes bibliográficos
Autor: Rangel Pardo, Francisco Manuel
Formato: tesis doctoral
Fecha de publicación:2016
País:España
Recursos:Universitat Politècnica de València (UPV)
Repositorio:RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Idioma:español
OAI Identifier:oai:riunet.upv.es:10251/67270
Acesso em linha:https://riunet.upv.es/handle/10251/67270
Access Level:acceso abierto
Palavra-chave:Author profiling
Age identification
Gender identification
Emograph
Language variety identification
Similar languages discrimination
Low dimensionality representation
Social media
Big data
LENGUAJES Y SISTEMAS INFORMATICOS
Descrição
Resumo:[EN] The possibility of knowing people traits on the basis of what they write is a field of growing interest named author profiling. To infer a user's gender, age, native language or personality traits, simply by analysing her texts, opens a wide range of possibilities from the point of view of forensics, security and marketing. Furthermore, social media proliferation, which allows for new communication models and human relations, strengthens this wide range of possibilities to bounds never seen before. Idiosyncrasy inherent to social media makes them a special environment of communication, where freedom of expression, informality and spontaneous generation of topics and trends, enhances the knowledge of the daily reality of people in their use of language. However, the same idiosyncrasy makes difficult, or extremely costly, the application of linguistic techniques. In this work we have proposed EmoGraph, a graph-based approach with the aim at modelling the way that users express their emotions, and the way they include them in their discourse, bearing in mind not only their frequency of occurrence, but also their position and relationship with other elements in the discourse. Our starting hypothesis is that users express themselves and their emotions differently depending on their age and gender, and besides, we think that this is independent on their language and social media where they write. We have collaborated in the creation of a common framework of evaluation at the PAN Lab of CLEF, generating resources that allowed us to verify our hypothesis achieving comparable and competitive results with the best ones obtained by other researchers on the field. In addition, we have investigated whether the expression of emotions would help to differentiate among users of different varieties of the same language, for example, Spanish from Spain, Mexican and Argentinian, or Portuguese from Portugal and Brazil. Our hypothesis is that the variation among languages is based more on lexical aspects, and we have corroborated it after comparing EmoGraph with representations based on word patterns, distributed representations and a representation that uses the whole vocabulary, but reducing its dimensionality to only 6 features per class, what is suitable for its application to big data environments such as social media.