UOBIT @ TAG-it: Exploring a Multi-faceted Representation for Profiling Age, Topic and Gender in Italian Texts

Abstract

English. This paper describes our system for participating in the TAG-it Author Profiling task at EVALITA 2020. The task aims to predict age and gender of blogs users from their posts, as the topic they wrote about. Our proposal combines learned representations by RNN at word and sentence levels, Transformer Neural Nets and hand-crafted stylistic features. All these representations are mixed and fed into a fully connected layer from a feed-forward neural network in order to make predictions for addressed subtasks. Experimental results show that our model achieves encouraging performance. The growing integration of social media with people’s daily live has made this medium a common environment for the deployment of technologies that allow the retrieval of useful information in the development of business activities, social outreach processes, forensic tasks, etc. That is because people frequently upload and share content in these media with various purposes such as socialization of points of view about some topic or promotion of personal business, etc. The analysis of textual information from such data, is one of the main reasons why researches become trending on the Natural Language Processing (NLP) field. However, the fact that this information varies greatly in terms of its format, even when it comes from the same person, besides textual sequences are unstructured information, make challenging the process of analyzing it automatically. Author Profiling (AP) task aims at discovering different marks or patterns (linguistic or not) from texts, that allow a user to be characterized in terms of Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). their age, gender, personality or any other demographic attribute. Many forums, due to the applicability of AP, share tasks directed to mining features that in general way, predict that valuable information. Those tasks commonly make special focus on popular languages such as English and Spanish. Nevertheless, other languages are explored on important forums too, that is the case of EVALITA 1, this one, promoting analysis of NLP tasks in the Italian language. Among the challenges from its last campaign EVALITA 2018 was the AP (in terms of gender) task GxG (Dell’Orletta and Nissim, 2018), exploring the gender-predicting issue. The analysis of age, gender and the topic a text is related with, are tasks well explored and the most approaches employ data representation based on stylistic features, n-gram representations and/or words embedding combined with Machine Learning (ML) methods like Support Vector Machine (SVM) and Random Forest (Pizarro, 2019). Also some authors by using Deep Learning (DL) models like Convolutional Neural Networks (CNN) and Long-Short Term Memory (LSTM) combined with stylistic features (Aragón and López-Monroy, 2018) (Bayot and Gonçalves, 2018) have yield encouraging performances. In this work we address precisely, the automatic detection of gender and age of the authors, besides the identification of the prevailing topic on textual information from blogs. Also, we describe our developed model for participating on TAG-it: Topic, Age and Gender prediction for Italian2 (Cimino A., 2020) task at EVALITA 2020 (Basile et al., 2020). Having in account the proved ability of DL http://www.evalita.it/ https://sites.google.com/view/