Abstract

Speaker: Sheharbano Khattak

Title: Characterization of Internet Censorship from Multiple Perspectives

Abstract: Censorship of online communications threatens principles of openness and freedom of information on which the Internet was founded. In the interest of transparency and accountability, and more broadly to develop scientific rigour in the field, we need methodologies to measure and characterize Internet censorship.  Such studies will not only help users make informed choices about information access, but also illuminate entities involved in or affected by censorship; informing the development of policy and enquiries into the ethics and legality of such practices. However, measurement of Internet censorship is more complex than typical communication network measurements because of the inherently adversarial and opaque landscape in which it operates. As details about mechanisms and targets of censorship are usually undisclosed, it is hard to define exactly what comprises censorship, and how it operates in different contexts. My research aims to help fill this gap by developing three methodologies, that are then applied to real-world datasets to characterize Internet censorship from multiple perspectives; uncovering entities involved in censorship and targets of censorship, and the effects of such practices on stakeholders. In this talk, I will provide an overview of the key contributions, results, and impact of my PhD thesis.

———

Speaker: Youmna Farag

Title: Convolutional Neural Networks for Automated Essay Assessment

Abstract: The task of assessing students’ essays has traditionally relied on human graders and their judgement of writing quality. There has been several attempts to build systems that can automate this task to make it more cost- and time-efficient. Most of these systems have heavily relied on handcrafted textual features to discriminate between well and poorly written essays. Since extracting such features is a daunting process, we propose to employ neural networks to the task of essay scoring. The networks operate on simple word vector representations, hence, avoiding expensive feature engineering. We, particularly, apply Convolutional Neural Networks (CNNs). CNNs have played a fundamental role in the recent breakthroughs in computer vision and object identification. The networks use image pixels and multiple levels of abstraction to effectively identify objects in images. We apply the same idea to predict essay scores by building networks that convolves over the different textual units: characters, words and sentences. Our results are promising and indicative of the ability of CNNs to evaluate writing quality.

———

Speaker: Anita Verő

Title: Questions in multi-modal semantics

Abstract: In multi-modal semantics we aim to ground meaning in perceptual input, usually using images and high performing deep visual representations, learned by convolutional neural networks. However, we still have many open questions about the sources and models we use, such as whether: 1) the choice of the network architecture affect performance, 2) the difference between search engines and manually annotated data sources has an impact, 3) the number of images for each word matter and 4) whether these findings extend to different languages? In my talk I will present our results investigating these issues. In the second part I will talk about my next question: whether multi-modality helps with the compositional representation of phrase and sentence meaning.