Master's Thesis Alexandre Mercier

Investigating data to text approaches to achieve diversity of generated marketing text in the music industry

Introduction & Motivation

The advent of large language models has completely redefined the generation of human-like text from structured data. In the context of digital platforms providing many similar but competing options, the main deciding factor are the descriptions of the goods or services. Therefore generating diverse and compelling marketing content plays a pivotal role in capturing user attention and fostering engagement. Unfortunately even LLMs still suffer from the problem of repetitions as shown in the paper "ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models" [1].

We are teaming up with startup Connactz, a platform aimed at connecting musicians and eventplanners to generate descriptions and press kits for artists as well as regional summaries. The thesis investigates the impact of different input data representations, training strategies, and model configurations on the diversity of the generated text while trying to keep the generated paragraphs fluent, relevant and correct.

Most studies so far have concentrated on reducing repetitions or controlling text diversity inside a single generated sample. In contrast, the aim of this study is to avoid similar structures and increase text diversity in between samples. By comparing and contrasting various approaches, the study aims to identify the most effective techniques for achieving diverse and engaging marketing content within the music industry context.

Research Questions

1.Which generative data-to-text approach yields the best overall results?

2.How can creativity/variety of generative models be controlled and evaluated? How unique is the generated content?

3.How can we compare and rank similar generated text?

4.Are generated texts as fluent and coherent as human written text?

5.Are current NLG techniques accurate enough for a production environment and in contact with users?

References

[1]    S. Jentzsch and K. Kersting, “ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models,” Jun. 2023. [Online]. Available: http://arxiv.org/pdf/2306.04563v1

TRANSLATE with x

English

Arabic	Hebrew	Polish
Bulgarian	Hindi	Portuguese
Catalan	Hmong Daw	Romanian
Chinese Simplified	Hungarian	Russian
Chinese Traditional	Indonesian	Slovak
Czech	Italian	Slovenian
Danish	Japanese	Spanish
Dutch	Klingon	Swedish
English	Korean	Thai
Estonian	Latvian	Turkish
Finnish	Lithuanian	Ukrainian
French	Malay	Urdu
German	Maltese	Vietnamese
Greek	Norwegian	Welsh
Haitian Creole	Persian

 

TRANSLATE with

COPY THE URL BELOW

EMBED THE SNIPPET BELOW IN YOUR SITE

 

Enable collaborative features and customize widget: Bing Webmaster Portal

 

"; langMenu.appendChild(origLangDiv); LanguageMenu.Init('LanguageMenu', LanguageMenu_keys, LanguageMenu_values, LanguageMenu_callback, LanguageMenu_popupid); window["LanguageMenu"] = LanguageMenu; clearInterval(intervalId); } }, 1); // ]]>