2020 Award Recipients
Segun Taofeek Aroyehun
Instituto Politécnico Nacional, Mexico
Dissertation title: Deep Learning Methods for Mining Social Media
As social media is nowadays at the core of human communication and interactions, it provides access to a huge amount of user-generated content, but at the same time is rife with harmful behavior. A problem both in efficiently using this communication opportunity and in addressing the challenges it creates is the creative and nuanced natural language expressions prevalent on social media. This research is directed towards natural language understanding in the social media domain by developing robust natural language processing and deep learning techniques. It aims to create end-user content moderation tools for preventing and countering harmful behaviors on social media, as well as applications that will drive public health safety in pharmacovigilance and disaster management. The research is conducted under the guidance of Professor Alexander Gelbukh at the Centro de Investigación en Computación (CIC) of the Instituto Politécnico Nacional (IPN), Mexico.
Universidad Nacional de Río Cuarto, Argentina
Dissertation title: Applying Learning Techniques to Oracle Synthesis and Analysis
The oracle problem refers to the challenge of distinguishing between the expected behavior of software and its actual behavior. Automatically deciding whether the software behavior is correct is in general very difficult, especially because executable software specifications, which can precisely capture expected behavior, are rarely found in practice. The absence of precise executable oracles represents a significant bottleneck that inhibits greater effectiveness in testing and bug finding, among other tasks. Thus, devising novel and effective mechanisms to distinguish valid from invalid software executions would have an enormous impact on the ability of automated analyses to detect program defects.
My research dissertation will focus on the application of learning techniques to tackle the oracle problem, in particular, around two different subproblems of the automated construction of software specifications:
- The translation of software specifications across different specification styles
- The synthesis of program specifications from program behaviors, and their analysis
Jéssica Soares dos Santos
Universidade Federal Fluminense, Brazil
Dissertation title: Mining Opinions in the Electoral Scenario Based on Transfer Learning
Analyzing opinions towards elections is a regular problem in society. The analysis of data extracted from social media has emerged as a faster and cheaper alternative to traditional election polls. However, social media data on this domain contain specific terms that evolve rapidly over time and it is impossible to manually annotate the huge amount of data during the short period of campaigns. We investigate how to boost and speed up the performance of opinion analysis tasks during elections by proposing a transfer learning approach that leverages curated datasets from several domains. To avoid introducing knowledge from the other domains that could end up disturbing the task, we propose to use similarity metrics to point out whether or not the dataset should be used. The same idea could be adopted to solve tasks of domains that undergo the same issues: short time for labeling and dynamic vocabulary.