Large language models (LLMs) have been applied to a wide range ofdata-to-text generation tasks, including tables, graphs, and time-seriesnumerical data-to-text settings. While research on generating prompts forstructured data such as tables and graphs is gaining momentum, in-depthinvestigations into prompting for time-series numerical data are lacking.Therefore, this study explores various input representations, includingsequences of tokens and structured formats such as HTML, LaTeX, andPython-style codes. In our experiments, we focus on the task of Market CommentGeneration, which involves taking a numerical sequence of stock prices as inputand generating a corresponding market comment. Contrary to our expectations,the results show that prompts resembling programming languages yield betteroutcomes, whereas those similar to natural languages and longer formats, suchas HTML and LaTeX, are less effective. Our findings offer insights intocreating effective prompts for tasks that generate text from numericalsequences.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)