This paper presents a text generation approach that involves copying and pasting text segments from an existing collection, resulting in better generation quality and comparable inference efficiency to autoregressive models. Domain adaptation and performance gains are also observed.
00:00 Section: 1 Introduction
03:09 Section: 2 Background: Neural Text Generation
05:40 Section: 3 Copy-Generator
07:59 Section: Ethical Consideration
11:14 Section: Context-Independent Token Embeddings
14:07 Section: 4 Experimental Setup
17:41 Section: 4.3 Automatic Evaluation Metrics
20:33 Section: Results
23:24 Section: Case Study
26:27 Section: Results
28:56 Section: Dense Retrieval
YouTube: @ArxivPapers
PODCASTS:
Apple Podcasts:
Spotify: