Conferència: "Project Angel: Technology to protect teenagers from hate speech and sexism", a càrrer de Vasudema Varma

El centre d’investigació PRHLT de la Universitat Politècnica de València, en el marc del projecte MISMIS (MISinformation and MIScommunication in Social Media, PGC2018-096212-B), en el qual participen la UPV, la UB i la UNED, i l'IP del qual és el professor amb PI Paolo Rosso, organitza una sèrie de xerrades sobre discursos d’odi i llenguatge tòxic en línia. En aquest projecte, per part de la UB, hi participa el grup de recerca CLiC, Centre de Llenguatge i Computació.

Dilluns 21 de desembre a les 11:00 h tindrà lloc la segona xerrada, amb el títol "Project Angel: Technology to protect teenagers from hate speech and sexism", a càrrec del professor Vasudema Varma (International Institute of Information Technology, Hyderabad).

Biografia: Vasudema Varma va ser degà de l'IIIT Hydebarad, on dirigeix l'Information Retrieval and Extraction Lab (iREL). Es va doctorar l'any 1996 per la Universitat de Hyderabad, i investiga àmbits com l'Arquitectura de Software; o l'Accés, Recuperació i Extracció de la Informació. 

En aquesta conferència també hi participaran dos dels seus estudiants: Harika Abburi i Sayar Gosh, que presentaran la seva recerca en les àrees de Classificació del Sexisme, i Generació de Text i Parla, respectivament. 

Abstract: "Project ANGEL" is an initiative aimed at utilizing technology, especially machine learning, for enhancing the well-being of teenagers, especially teenage girls. In many cases, the digital lives led by young people reflect various kinds of troubles afflicting them online and offline. We intend to develop a cohesive technology stack through multi-disciplinary research for helping teen-age people in an empathetic, proactive manner. In this talk, we focus more on a novel neural framework for classifying sexism and misogyny that can combine text representations obtained using models such as BERT with distributional and linguistic word embeddings using a flexible architecture involving recurrent components and optional convolutional ones. Further, we leverage unlabeled accounts of sexism to infuse domain-specific elements into our framework. We investigate more fine-grained, multi-label classification of accounts (reports) of sexism.

Moreover, we propose a multi-task approach for fine-grained multi-label sexism classification that leverages several supporting tasks without incurring any manual labeling cost. Unlabeled accounts of sexism are utilized through unsupervised learning to help construct our multi-task setup. We also devise objective functions that exploit label correlations in the training data explicitly. We present some of our work in progress and some of our future plans in terms of identifying the hate speech intervention points and generating appropriate response to counter the hate speech and sexism comments. We share some of our learnings from the recent shared tasks on hate speech detection. Specifically, we look at the usage of Transformer models pretrained on multilingual data. We comment on the effectiveness of leveraging features from publicly available toxicity-level detection modules. The idea of utilizing a task adaptive pretraining approach for hate speech and fake news detection in English and Hindi is touched upon. We discuss some of the recent advances in online hate speech intervention and throw light on our future objectives in the space of response generation.