This is an online seminar. Registration is required.
【 Natural Language Understanding Team】
【Date】2023/December/4(Mon) 17:30-18:30(JST)
【Speaker】Zhang Ying ,Tokyo Institute of Technology, Doctoral course in Information and Communications Engineering
Title: Addressing Text Degeneration of Discriminative Models with Re-ranking Methods
Abstract: Despite challenges such as unbalanced data distribution and limited training data,
discriminative models are often preferred over generative models in natural language
processing tasks owing to their direct computation and proven success in various applications.
However, recent studies have highlighted two significant issues: text degeneration and
exposure bias, that arise when using discriminative models.
The primary objective of this study is to improve the prediction results from discriminative
models by presenting two re-ranking-based methods to counter the issues of text
degeneration and exposure bias. The aim of this study is to propose general ideas that can
be easily implemented in natural language processing tasks, including grammatical error
correction and discourse parsing, without requiring complex modifications to the existing
discriminative model structure or excessive computational resources.