This Focus on Research Seminar has been cancelled.

When: Monday, March 23, 2020
Where: PGH 563
Time: 11:00 AM

CANCELLED

Focus on Research (FoR) is an opportunity for any COSC Ph.D. student to discuss a research project (with or without preliminary results), a conference dry run, or any research topic of interest to present to an audience of peers and faculty. It is a great avenue for Ph.D. students to practice presentation skills in front of a larger and broader audience.

On the Usefulness of Personality Traits in Opinion-Oriented Tasks

Marjan Hosseinia, Ph.D. Student

Abstract:

We use a deep bidirectional transformer to extract Myers-Briggs personality types from user-generated data in multi-label classification mode. Our dataset is large and made up of three available personality data of various social media platforms including Reddit, Twitter, and a personality forum. We infer personality information from our transformer-based model and investigate if it can be useful for downstream opinion-oriented text classification tasks. Experimental evidence shows the effectiveness of the pre-trained model on personality data in stance detection, authorship verification, and sentiment analysis.

Marjan Hosseinia is a Ph.D. candidate at University of Houston. She works under the supervision of Dr. Arjun Mukherjee. Her research interests are natural language processing and opinion mining.

Explainable Generative Adversarial Neural Networks

Hadi Mansourifar, Ph.D. Student

Abstract:

Generative Adversarial Networks (GANs) are being increasingly used in many tasks from data augmentation to anomaly detection, due to their capabilities of automatically generating highly-realistic synthetic images in an adversarial mode mostly based on Convolutional Neural Networks (CNN). However, GANs currently represent black box models because of a variety of reasons: It’s not possible to detect the mode collapse and vanishing generator gradients as two major problems of the GANs. Besides, training loss does not reveal anything about the quality of generated fake instances. In this research, we try to make the GANs explainable via proposing a new measure to evaluate the similarity of fake-real data distribution. The proposed measure makes it possible to observe the mode collapse and vanishing generator gradient. Furthermore, it enables us to track the impact of discriminator power in a training phase.

Bio:

Hadi Mansourifar is a fourth-year Ph.D. student who works under the supervision of Dr. Larry Shi.

Claim Verification under the Positive-Unlabeled Setting

Fan Yang, Ph.D. Student

Abstract:

We extend claim verification to the context of positive-unlabeled (PU) learning. Existing works assume the truth and the falsity of the claims are known for training, and forms the task as a supervised learning problem. However, this assumption underestimates the difficulty of collecting false claims; we argue that claim verification is more challenging in the absence of negative labels. We consider a more practical setting, where only a comparatively small number of true claims are labeled and more claims remain unlabeled. Thus, we formulate the claim verification task as a PU learning problem. We decouple learning representation of claim-evidence pairs from PU learning and adopt a pre-trained universal language model to encode claim-evidence pairs. We further propose to use the generative adversarial network (GAN) to capture the latent alignment between encoded claim-evidence pairs and the truthfulness. We leverage the verification as part of the GAN by extending previous GAN-based PU learning. We show that the proposed model achieves the best performance with a small amount of labeled data and is robust to the truthfulness prior to estimation. We conduct a thorough analysis of the model selection. The proposed approach performs the best under two practical scenarios: 1) the unlabeled data is more than the labeled data, and 2) the unlabeled positive data is more than the unlabeled negative data.

Bio:

Fan Yang is a fifth-year Ph.D. student, advised by Dr. Arjun Mukherjee. Fan is interested in deep learning and natural language understanding, with a particular focus on detecting misleading information.